Add files using upload-large-folder tool

.gitattributes

@@ -41,3 +41,18 @@ dino_model/facebookresearch_dino_main/.github/dino.gif filter=lfs diff=lfs merge
 dreamsim_model/facebookresearch_dino_main/.github/attention_maps.png filter=lfs diff=lfs merge=lfs -text
 dreamsim_model/facebookresearch_co-tracker_main/notebooks/demo.ipynb filter=lfs diff=lfs merge=lfs -text
 torch_hub/facebookresearch_dino_main/.github/dino.gif filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_dinov2_main/docs/ChannelAdaptiveDINO.png filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_dinov2_main/docs/Cell-DINO.png filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/teddy.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/paragliding.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/assets/apple.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_dino_main/.github/attention_maps.png filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/notebooks/demo.ipynb filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/bear.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/cat.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/apple.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/backpack.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/pillow.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/gradio_demo/videos/paragliding-launch.mp4 filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/assets/teaser.png filter=lfs diff=lfs merge=lfs -text
+torch_hub/facebookresearch_co-tracker_main/assets/bmx-bumps.gif filter=lfs diff=lfs merge=lfs -text

torch_hub/checkpoints/alexnet-owt-7be5be79.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7be5be791159472b1fbf3c69796f7cb30dca7ad8466c2df70058c37116cdee02
+size 244408911

torch_hub/checkpoints/convnext_tiny_1k_224_ema.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14f3164e3ea6ac32ab3f574f528ce817696c9176fad4221e0a77a905a7360595
+size 114414741

torch_hub/checkpoints/cotracker2.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:362f5274376d610dc987b6daf2c2fefe63e06e1835f4ec1a10d0a15c5a4eef4f
+size 204396415

torch_hub/checkpoints/dino_vitbase16_pretrain.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bf34ad0f424b9029b593e8dc3ed553bf26e88bcba0d32bf3e62a6209cb64c85e
+size 343242485

torch_hub/checkpoints/dinov2_vitb14_reg4_pretrain.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:73182a088cf94833c94b1666d1c99e02fe87e2007bff57b564fb6206e25dba71
+size 346393545

torch_hub/checkpoints/pt_inception-2015-12-05-6726825d.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6726825d0af5f729cebd5821db510b11b1cfad8faad88a03f1befd49fb9129b2
+size 95628359

torch_hub/checkpoints/vgg16-397923af.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:397923af8e79cdbb6a7127f12361acd7a2f83e06b05044ddf496e83de57a5bf0
+size 553433881

torch_hub/checkpoints/weights-inception-2015-12-05-6726825d.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6726825d0af5f729cebd5821db510b11b1cfad8faad88a03f1befd49fb9129b2
+size 95628359

torch_hub/facebookresearch_co-tracker_main/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <opensource-conduct@fb.com>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

torch_hub/facebookresearch_co-tracker_main/CONTRIBUTING.md

@@ -0,0 +1,28 @@
+# CoTracker
+We want to make contributing to this project as easy and transparent as possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've changed APIs, update the documentation.
+3. Make sure your code lints.
+4. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Meta's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Meta has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to CoTracker, you agree that your contributions will be licensed
+under the LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/LICENSE.md

@@ -0,0 +1,399 @@
+Attribution-NonCommercial 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+	wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More_considerations
+     for the public: 
+	wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial 4.0 International Public
+License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial 4.0 International Public License ("Public
+License"). To the extent this Public License may be interpreted as a
+contract, You are granted the Licensed Rights in consideration of Your
+acceptance of these terms and conditions, and the Licensor grants You
+such rights in consideration of benefits the Licensor receives from
+making the Licensed Material available under these terms and
+conditions.
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+  d. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  e. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  f. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  g. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  h. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  i. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  j. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  k. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  l. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+       4. If You Share Adapted Material You produce, the Adapter's
+          License You apply must not prevent recipients of the Adapted
+          Material from complying with this Public License.
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material; and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

torch_hub/facebookresearch_co-tracker_main/README.md

@@ -0,0 +1,360 @@
+# CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos
+
+**[Meta AI Research, GenAI](https://ai.facebook.com/research/)**; **[University of Oxford, VGG](https://www.robots.ox.ac.uk/~vgg/)**
+
+[Nikita Karaev](https://nikitakaraevv.github.io/), [Iurii Makarov](https://linkedin.com/in/lvoursl), [Jianyuan Wang](https://jytime.github.io/), [Ignacio Rocco](https://www.irocco.info/), [Benjamin Graham](https://ai.facebook.com/people/benjamin-graham/), [Natalia Neverova](https://nneverova.github.io/), [Andrea Vedaldi](https://www.robots.ox.ac.uk/~vedaldi/), [Christian Rupprecht](https://chrirupp.github.io/)
+
+### [Project Page](https://cotracker3.github.io/) | [Paper #1](https://arxiv.org/abs/2307.07635) | [Paper #2](https://arxiv.org/abs/2410.11831) |  [X Thread](https://twitter.com/n_karaev/status/1742638906355470772) | [BibTeX](#citing-cotracker)
+
+<a target="_blank" href="https://colab.research.google.com/github/facebookresearch/co-tracker/blob/main/notebooks/demo.ipynb">
+  <img src="https://colab.research.google.com/assets/colab-badge.svg" alt="Open In Colab"/>
+</a>
+<a href="https://huggingface.co/spaces/facebook/cotracker">
+  <img alt="Spaces" src="https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue">
+</a>
+
+<img width="1100" src="./assets/teaser.png" />
+
+**CoTracker** is a fast transformer-based model that can track any point in a video. It brings to tracking some of the benefits of Optical Flow.
+
+CoTracker can track:
+
+- **Any pixel** in a video
+- A **quasi-dense** set of pixels together
+- Points can be manually selected or sampled on a grid in any video frame
+
+Try these tracking modes for yourself with our [Colab demo](https://colab.research.google.com/github/facebookresearch/co-tracker/blob/master/notebooks/demo.ipynb) or in the [Hugging Face Space 🤗](https://huggingface.co/spaces/facebook/cotracker).
+
+**Updates:**
+
+- [January 21, 2025] 📦 Kubric Dataset used for CoTracker3 now available! This dataset contains **6,000 high-resolution sequences** (512×512px, 120 frames) with slight camera motion, rendered using the Kubric engine. Check it out on [Hugging Face Dataset](https://huggingface.co/datasets/facebook/CoTracker3_Kubric).
+
+- [October 15, 2024] 📣 We're releasing CoTracker3! State-of-the-art point tracking with a lightweight architecture trained with 1000x less data than previous top-performing models. Code for baseline models and the pseudo-labeling pipeline are available in the repo, as well as model checkpoints. Check out our [paper](https://arxiv.org/abs/2410.11831) for more details.
+
+- [September 25, 2024]  CoTracker2.1 is now available! This model has better performance on TAP-Vid benchmarks and follows the architecture of the original CoTracker. Try it out!
+
+- [June 14, 2024]  We have released the code for [VGGSfM](https://github.com/facebookresearch/vggsfm), a model for recovering camera poses and 3D structure from any image sequences based on point tracking! VGGSfM is the first fully differentiable SfM framework that unlocks scalability and outperforms conventional SfM methods on standard benchmarks. 
+
+- [December 27, 2023]  CoTracker2 is now available! It can now track many more (up to **265*265**!) points jointly and it has a cleaner and more memory-efficient implementation. It also supports online processing. See the [updated paper](https://arxiv.org/abs/2307.07635) for more details. The old version remains available [here](https://github.com/facebookresearch/co-tracker/tree/8d364031971f6b3efec945dd15c468a183e58212).
+
+- [September 5, 2023] You can now run our Gradio demo [locally](./gradio_demo/app.py).
+
+## Quick start
+The easiest way to use CoTracker is to load a pretrained model from `torch.hub`:
+
+### Offline mode: 
+```pip install imageio[ffmpeg]```, then:
+```python
+import torch
+# Download the video
+url = 'https://github.com/facebookresearch/co-tracker/raw/refs/heads/main/assets/apple.mp4'
+
+import imageio.v3 as iio
+frames = iio.imread(url, plugin="FFMPEG")  # plugin="pyav"
+
+device = 'cuda'
+grid_size = 10
+video = torch.tensor(frames).permute(0, 3, 1, 2)[None].float().to(device)  # B T C H W
+
+# Run Offline CoTracker:
+cotracker = torch.hub.load("facebookresearch/co-tracker", "cotracker3_offline").to(device)
+pred_tracks, pred_visibility = cotracker(video, grid_size=grid_size) # B T N 2,  B T N 1
+```
+### Online mode: 
+```python
+cotracker = torch.hub.load("facebookresearch/co-tracker", "cotracker3_online").to(device)
+
+# Run Online CoTracker, the same model with a different API:
+# Initialize online processing
+cotracker(video_chunk=video, is_first_step=True, grid_size=grid_size)  
+
+# Process the video
+for ind in range(0, video.shape[1] - cotracker.step, cotracker.step):
+    pred_tracks, pred_visibility = cotracker(
+        video_chunk=video[:, ind : ind + cotracker.step * 2]
+    )  # B T N 2,  B T N 1
+```
+Online processing is more memory-efficient and allows for the processing of longer videos. However, in the example provided above, the video length is known! See [the online demo](./online_demo.py) for an example of tracking from an online stream with an unknown video length.
+
+### Visualize predicted tracks: 
+After [installing](#installation-instructions) CoTracker, you can visualize tracks with:
+```python
+from cotracker.utils.visualizer import Visualizer
+
+vis = Visualizer(save_dir="./saved_videos", pad_value=120, linewidth=3)
+vis.visualize(video, pred_tracks, pred_visibility)
+```
+
+We offer a number of other ways to interact with CoTracker:
+
+1. Interactive Gradio demo:
+   - A demo is available in the [`facebook/cotracker` Hugging Face Space 🤗](https://huggingface.co/spaces/facebook/cotracker).
+   - You can use the gradio demo locally by running [`python -m gradio_demo.app`](./gradio_demo/app.py) after installing the required packages: `pip install -r gradio_demo/requirements.txt`.
+2. Jupyter notebook:
+   - You can run the notebook in
+   [Google Colab](https://colab.research.google.com/github/facebookresearch/co-tracker/blob/master/notebooks/demo.ipynb).
+   - Or explore the notebook located at [`notebooks/demo.ipynb`](./notebooks/demo.ipynb). 
+2. You can [install](#installation-instructions) CoTracker _locally_ and then:
+   - Run an *offline* demo with 10 ⨉ 10 points sampled on a grid on the first frame of a video (results will be saved to `./saved_videos/demo.mp4`)):
+
+     ```bash
+     python demo.py --grid_size 10
+     ```
+    - Run an *online* demo:
+
+      ```bash
+      python online_demo.py
+      ```
+
+A GPU is strongly recommended for using CoTracker locally.
+
+<img width="500" src="./assets/bmx-bumps.gif" />
+
+
+## Installation Instructions
+You can use a Pretrained Model via PyTorch Hub, as described above, or install CoTracker from this GitHub repo.
+This is the best way if you need to run our local demo or evaluate/train CoTracker.
+
+Ensure you have both _PyTorch_ and _TorchVision_ installed on your system. Follow the instructions [here](https://pytorch.org/get-started/locally/) for the installation.
+We strongly recommend installing both PyTorch and TorchVision with CUDA support, although for small tasks CoTracker can be run on CPU.
+
+
+
+
+### Install a Development Version
+
+```bash
+git clone https://github.com/facebookresearch/co-tracker
+cd co-tracker
+pip install -e .
+pip install matplotlib flow_vis tqdm tensorboard
+```
+
+You can manually download all CoTracker3 checkpoints (baseline and scaled models, as well as single and sliding window architectures) from the links below and place them in the `checkpoints` folder as follows:
+
+```bash
+mkdir -p checkpoints
+cd checkpoints
+# download the online (multi window) model
+wget https://huggingface.co/facebook/cotracker3/resolve/main/scaled_online.pth
+# download the offline (single window) model
+wget https://huggingface.co/facebook/cotracker3/resolve/main/scaled_offline.pth
+cd ..
+```
+You can also download CoTracker3 checkpoints trained only on Kubric:
+```bash
+# download the online (sliding window) model
+wget https://huggingface.co/facebook/cotracker3/resolve/main/baseline_online.pth
+# download the offline (single window) model
+wget https://huggingface.co/facebook/cotracker3/resolve/main/baseline_offline.pth
+```
+For old checkpoints, see [this section](#previous-version).
+
+## Evaluation
+
+To reproduce the results presented in the paper, download the following datasets:
+
+- [TAP-Vid](https://github.com/deepmind/tapnet)
+- [Dynamic Replica](https://dynamic-stereo.github.io/)
+
+And install the necessary dependencies:
+
+```bash
+pip install hydra-core==1.1.0 mediapy
+```
+
+Then, execute the following command to evaluate the online model on TAP-Vid DAVIS:
+
+```bash
+python ./cotracker/evaluation/evaluate.py --config-name eval_tapvid_davis_first exp_dir=./eval_outputs dataset_root=your/tapvid/path
+```
+And the offline model:
+```bash
+python ./cotracker/evaluation/evaluate.py --config-name eval_tapvid_davis_first exp_dir=./eval_outputs dataset_root=/fsx-repligen/shared/datasets/tapvid offline_model=True window_len=60 checkpoint=./checkpoints/scaled_offline.pth
+```
+We run evaluations jointly on all the target points at a time for faster inference. With such evaluations, the numbers are similar to those presented in the paper. If you want to reproduce the exact numbers from the paper, add the flag `single_point=True`. 
+
+These are the numbers that you should be able to reproduce using the released checkpoint and the current version of the codebase:
+|  | Kinetics, $\delta_\text{avg}^\text{vis}$ | DAVIS, $\delta_\text{avg}^\text{vis}$ |  RoboTAP, $\delta_\text{avg}^\text{vis}$ | RGB-S, $\delta_\text{avg}^\text{vis}$| 
+| :---: |:---: | :---: | :---: | :---: |
+| CoTracker2, 27.12.23 | 61.8 | 74.6 | 69.6 | 73.4 | 
+| CoTracker2.1, 25.09.24 | 63 | 76.1 | 70.6 | 79.6 | 
+| CoTracker3 offline, 15.10.24 | 67.8 | **76.9** | 78.0 | **85.0** | 
+| CoTracker3 online, 15.10.24 | **68.3** | 76.7 | **78.8** | 82.7  | 
+
+
+## Training
+
+### Baseline
+To train the CoTracker as described in our paper, you first need to generate annotations for [Google Kubric](https://github.com/google-research/kubric) MOVI-f dataset.
+Instructions for annotation generation can be found [here](https://github.com/deepmind/tapnet).
+You can also find a discussion on dataset generation in [this issue](https://github.com/facebookresearch/co-tracker/issues/8).
+
+Once you have the annotated dataset, you need to make sure you followed the steps for evaluation setup and install the training dependencies:
+
+```bash
+pip install pip==24.0
+pip install pytorch_lightning==1.6.0 tensorboard opencv-python
+```
+
+Now you can launch training on Kubric.
+Our model was trained for 50000 iterations on 32 GPUs (4 nodes with 8 GPUs). 
+Modify _dataset_root_ and _ckpt_path_ accordingly before running this command. For training on 4 nodes, add `--num_nodes 4`. 
+
+Here is an example of how to launch training of the online model on Kubric:
+```bash
+ python train_on_kubric.py --batch_size 1 --num_steps 50000 \
+ --ckpt_path ./ --model_name cotracker_three --save_freq 200 --sequence_len 64 \
+  --eval_datasets tapvid_davis_first tapvid_stacking --traj_per_sample 384 \
+  --sliding_window_len 16 --train_datasets kubric --save_every_n_epoch 5 \
+  --evaluate_every_n_epoch 5 --model_stride 4 --dataset_root ${path_to_your_dataset} \
+   --num_nodes 4 --num_virtual_tracks 64 --mixed_precision --corr_radius 3 \ 
+   --wdecay 0.0005 --linear_layer_for_vis_conf --validate_at_start --add_huber_loss
+```
+
+Training the offline model on Kubric:
+```bash
+python train_on_kubric.py --batch_size 1 --num_steps 50000 \
+ --ckpt_path ./ --model_name cotracker_three --save_freq 200 --sequence_len 60 \
+ --eval_datasets tapvid_davis_first tapvid_stacking --traj_per_sample 512 \
+ --sliding_window_len 60 --train_datasets kubric --save_every_n_epoch 5 \
+ --evaluate_every_n_epoch 5 --model_stride 4 --dataset_root ${path_to_your_dataset} \
+ --num_nodes 4 --num_virtual_tracks 64 --mixed_precision --offline_model \
+ --random_frame_rate --query_sampling_method random --corr_radius 3 \
+ --wdecay 0.0005 --random_seq_len --linear_layer_for_vis_conf \
+ --validate_at_start --add_huber_loss
+```
+
+### Fine-tuning with pseudo labels
+In order to launch training with pseudo-labelling, you need to collect your own dataset of real videos. There is a sample class available in [`cotracker/datasets/real_dataset.py`](./cotracker/datasets/real_dataset.py) with keyword-based filtering that we used for training. Your class should implement loading a video and storing it in the `CoTrackerData` class as a field, while pseudo labels will be generated in `train_on_real_data.py`.
+
+You should have an existing Kubric-trained model for fine-tuning with pseudo labels. Here is an example of how you can launch fine-tuning of the online model:
+```bash
+python ./train_on_real_data.py --batch_size 1 --num_steps 15000 \
+ --ckpt_path ./ --model_name cotracker_three --save_freq 200 --sequence_len 64 \
+ --eval_datasets tapvid_stacking tapvid_davis_first --traj_per_sample 384 \
+ --save_every_n_epoch 15 --evaluate_every_n_epoch 15 --model_stride 4 \
+ --dataset_root ${path_to_your_dataset} --num_nodes 4 --real_data_splits 0 \
+ --num_virtual_tracks 64 --mixed_precision --random_frame_rate \
+ --restore_ckpt ./checkpoints/baseline_online.pth \
+ --lr 0.00005 --real_data_filter_sift --validate_at_start \
+ --sliding_window_len 16 --limit_samples 15000
+
+```
+And the offline model:
+```bash
+python train_on_real_data.py --batch_size 1 --num_steps 15000 \
+ --ckpt_path ./ --model_name cotracker_three --save_freq 200 --sequence_len 80 \
+ --eval_datasets tapvid_stacking tapvid_davis_first --traj_per_sample 384 --save_every_n_epoch 15 \
+ --evaluate_every_n_epoch 15 --model_stride 4 --dataset_root ${path_to_your_dataset} \
+  --num_nodes 4 --real_data_splits 0 --num_virtual_tracks 64 --mixed_precision \
+  --random_frame_rate --restore_ckpt ./checkpoints/baseline_offline.pth --lr 0.00005 \
+  --real_data_filter_sift --validate_at_start --offline_model --limit_samples 15000
+```
+
+
+
+## Development
+
+### Building the documentation
+
+To build CoTracker documentation, first install the dependencies:
+
+```bash
+pip install sphinx
+pip install sphinxcontrib-bibtex
+```
+
+Then you can use this command to generate the documentation in the `docs/_build/html` folder:
+
+```bash
+make -C docs html
+```
+
+
+## Previous versions
+### CoTracker v2
+You could use CoTracker v2 with torch.hub in both offline and online modes.
+#### Offline mode: 
+```pip install imageio[ffmpeg]```, then:
+```python
+import torch
+# Download the video
+url = 'https://github.com/facebookresearch/co-tracker/blob/main/assets/apple.mp4'
+
+import imageio.v3 as iio
+frames = iio.imread(url, plugin="FFMPEG")  # plugin="pyav"
+
+device = 'cuda'
+grid_size = 10
+video = torch.tensor(frames).permute(0, 3, 1, 2)[None].float().to(device)  # B T C H W
+
+# Run Offline CoTracker:
+cotracker = torch.hub.load("facebookresearch/co-tracker", "cotracker2").to(device)
+pred_tracks, pred_visibility = cotracker(video, grid_size=grid_size) # B T N 2,  B T N 1
+```
+#### Online mode: 
+```python
+cotracker = torch.hub.load("facebookresearch/co-tracker", "cotracker2_online").to(device)
+
+# Run Online CoTracker, the same model with a different API:
+# Initialize online processing
+cotracker(video_chunk=video, is_first_step=True, grid_size=grid_size)  
+
+# Process the video
+for ind in range(0, video.shape[1] - cotracker.step, cotracker.step):
+    pred_tracks, pred_visibility = cotracker(
+        video_chunk=video[:, ind : ind + cotracker.step * 2]
+    )  # B T N 2,  B T N 1
+```
+
+Checkpoint for v2 could be downloaded with the following command:
+```bash
+wget https://huggingface.co/facebook/cotracker/resolve/main/cotracker2.pth
+```
+
+### CoTracker v1
+It is directly available via pytorch hub:
+```python
+import torch
+import einops
+import timm
+import tqdm
+
+cotracker = torch.hub.load("facebookresearch/co-tracker:v1.0", "cotracker_w8")
+```
+The old version of the code is available [here](https://github.com/facebookresearch/co-tracker/tree/8d364031971f6b3efec945dd15c468a183e58212).
+You can also download the corresponding checkpoints:
+```bash
+wget https://dl.fbaipublicfiles.com/cotracker/cotracker_stride_4_wind_8.pth
+wget https://dl.fbaipublicfiles.com/cotracker/cotracker_stride_4_wind_12.pth
+wget https://dl.fbaipublicfiles.com/cotracker/cotracker_stride_8_wind_16.pth
+```
+
+## License
+
+The majority of CoTracker is licensed under CC-BY-NC, however portions of the project are available under separate license terms: Particle Video Revisited is licensed under the MIT license, TAP-Vid and LocoTrack are licensed under the Apache 2.0 license.
+
+## Acknowledgments
+
+We would like to thank [PIPs](https://github.com/aharley/pips), [TAP-Vid](https://github.com/deepmind/tapnet), [LocoTrack](https://github.com/cvlab-kaist/locotrack) for publicly releasing their code and data. We also want to thank [Luke Melas-Kyriazi](https://lukemelas.github.io/) for proofreading the paper, [Jianyuan Wang](https://jytime.github.io/), [Roman Shapovalov](https://shapovalov.ro/) and [Adam W. Harley](https://adamharley.com/) for the insightful discussions.
+
+## Citing CoTracker
+
+If you find our repository useful, please consider giving it a star ⭐ and citing our research papers in your work:
+```bibtex
+@inproceedings{karaev23cotracker,
+  title     = {CoTracker: It is Better to Track Together},
+  author    = {Nikita Karaev and Ignacio Rocco and Benjamin Graham and Natalia Neverova and Andrea Vedaldi and Christian Rupprecht},
+  booktitle = {Proc. {ECCV}},
+  year      = {2024}
+}
+```
+```bibtex
+@inproceedings{karaev24cotracker3,
+  title     = {CoTracker3: Simpler and Better Point Tracking by Pseudo-Labelling Real Videos},
+  author    = {Nikita Karaev and Iurii Makarov and Jianyuan Wang and Natalia Neverova and Andrea Vedaldi and Christian Rupprecht},
+  booktitle = {Proc. {arXiv:2410.11831}},
+  year      = {2024}
+}
+```

torch_hub/facebookresearch_co-tracker_main/assets/apple.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c7f48c5cfb1479e1dbc1df2373d5cad4f55c198bbdb379da0ece10087971542a
+size 1219872

torch_hub/facebookresearch_co-tracker_main/cotracker/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/dataclass_utils.py

@@ -0,0 +1,168 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import json
+import dataclasses
+import numpy as np
+from dataclasses import Field, MISSING
+from typing import IO, TypeVar, Type, get_args, get_origin, Union, Any, Tuple
+
+_X = TypeVar("_X")
+
+
+def load_dataclass(f: IO, cls: Type[_X], binary: bool = False) -> _X:
+    """
+    Loads to a @dataclass or collection hierarchy including dataclasses
+    from a json recursively.
+    Call it like load_dataclass(f, typing.List[FrameAnnotationAnnotation]).
+    raises KeyError if json has keys not mapping to the dataclass fields.
+
+    Args:
+        f: Either a path to a file, or a file opened for writing.
+        cls: The class of the loaded dataclass.
+        binary: Set to True if `f` is a file handle, else False.
+    """
+    if binary:
+        asdict = json.loads(f.read().decode("utf8"))
+    else:
+        asdict = json.load(f)
+
+    # in the list case, run a faster "vectorized" version
+    cls = get_args(cls)[0]
+    res = list(_dataclass_list_from_dict_list(asdict, cls))
+
+    return res
+
+
+def _resolve_optional(type_: Any) -> Tuple[bool, Any]:
+    """Check whether `type_` is equivalent to `typing.Optional[T]` for some T."""
+    if get_origin(type_) is Union:
+        args = get_args(type_)
+        if len(args) == 2 and args[1] == type(None):  # noqa E721
+            return True, args[0]
+    if type_ is Any:
+        return True, Any
+
+    return False, type_
+
+
+def _unwrap_type(tp):
+    # strips Optional wrapper, if any
+    if get_origin(tp) is Union:
+        args = get_args(tp)
+        if len(args) == 2 and any(a is type(None) for a in args):  # noqa: E721
+            # this is typing.Optional
+            return args[0] if args[1] is type(None) else args[1]  # noqa: E721
+    return tp
+
+
+def _get_dataclass_field_default(field: Field) -> Any:
+    if field.default_factory is not MISSING:
+        # pyre-fixme[29]: `Union[dataclasses._MISSING_TYPE,
+        #  dataclasses._DefaultFactory[typing.Any]]` is not a function.
+        return field.default_factory()
+    elif field.default is not MISSING:
+        return field.default
+    else:
+        return None
+
+
+def _dataclass_list_from_dict_list(dlist, typeannot):
+    """
+    Vectorised version of `_dataclass_from_dict`.
+    The output should be equivalent to
+    `[_dataclass_from_dict(d, typeannot) for d in dlist]`.
+
+    Args:
+        dlist: list of objects to convert.
+        typeannot: type of each of those objects.
+    Returns:
+        iterator or list over converted objects of the same length as `dlist`.
+
+    Raises:
+        ValueError: it assumes the objects have None's in consistent places across
+            objects, otherwise it would ignore some values. This generally holds for
+            auto-generated annotations, but otherwise use `_dataclass_from_dict`.
+    """
+
+    cls = get_origin(typeannot) or typeannot
+
+    if typeannot is Any:
+        return dlist
+    if all(obj is None for obj in dlist):  # 1st recursion base: all None nodes
+        return dlist
+    if any(obj is None for obj in dlist):
+        # filter out Nones and recurse on the resulting list
+        idx_notnone = [(i, obj) for i, obj in enumerate(dlist) if obj is not None]
+        idx, notnone = zip(*idx_notnone)
+        converted = _dataclass_list_from_dict_list(notnone, typeannot)
+        res = [None] * len(dlist)
+        for i, obj in zip(idx, converted):
+            res[i] = obj
+        return res
+
+    is_optional, contained_type = _resolve_optional(typeannot)
+    if is_optional:
+        return _dataclass_list_from_dict_list(dlist, contained_type)
+
+    # otherwise, we dispatch by the type of the provided annotation to convert to
+    if issubclass(cls, tuple) and hasattr(cls, "_fields"):  # namedtuple
+        # For namedtuple, call the function recursively on the lists of corresponding keys
+        types = cls.__annotations__.values()
+        dlist_T = zip(*dlist)
+        res_T = [
+            _dataclass_list_from_dict_list(key_list, tp)
+            for key_list, tp in zip(dlist_T, types)
+        ]
+        return [cls(*converted_as_tuple) for converted_as_tuple in zip(*res_T)]
+    elif issubclass(cls, (list, tuple)):
+        # For list/tuple, call the function recursively on the lists of corresponding positions
+        types = get_args(typeannot)
+        if len(types) == 1:  # probably List; replicate for all items
+            types = types * len(dlist[0])
+        dlist_T = zip(*dlist)
+        res_T = (
+            _dataclass_list_from_dict_list(pos_list, tp)
+            for pos_list, tp in zip(dlist_T, types)
+        )
+        if issubclass(cls, tuple):
+            return list(zip(*res_T))
+        else:
+            return [cls(converted_as_tuple) for converted_as_tuple in zip(*res_T)]
+    elif issubclass(cls, dict):
+        # For the dictionary, call the function recursively on concatenated keys and vertices
+        key_t, val_t = get_args(typeannot)
+        all_keys_res = _dataclass_list_from_dict_list(
+            [k for obj in dlist for k in obj.keys()], key_t
+        )
+        all_vals_res = _dataclass_list_from_dict_list(
+            [k for obj in dlist for k in obj.values()], val_t
+        )
+        indices = np.cumsum([len(obj) for obj in dlist])
+        assert indices[-1] == len(all_keys_res)
+
+        keys = np.split(list(all_keys_res), indices[:-1])
+        all_vals_res_iter = iter(all_vals_res)
+        return [cls(zip(k, all_vals_res_iter)) for k in keys]
+    elif not dataclasses.is_dataclass(typeannot):
+        return dlist
+
+    # dataclass node: 2nd recursion base; call the function recursively on the lists
+    # of the corresponding fields
+    assert dataclasses.is_dataclass(cls)
+    fieldtypes = {
+        f.name: (_unwrap_type(f.type), _get_dataclass_field_default(f))
+        for f in dataclasses.fields(typeannot)
+    }
+
+    # NOTE the default object is shared here
+    key_lists = (
+        _dataclass_list_from_dict_list([obj.get(k, default) for obj in dlist], type_)
+        for k, (type_, default) in fieldtypes.items()
+    )
+    transposed = zip(*key_lists)
+    return [cls(*vals_as_tuple) for vals_as_tuple in transposed]

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/dr_dataset.py

@@ -0,0 +1,168 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import os
+import gzip
+import torch
+import numpy as np
+import torch.utils.data as data
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import List, Optional, Any, Dict, Tuple
+
+from cotracker.datasets.utils import CoTrackerData
+from cotracker.datasets.dataclass_utils import load_dataclass
+
+
+@dataclass
+class ImageAnnotation:
+    # path to jpg file, relative w.r.t. dataset_root
+    path: str
+    # H x W
+    size: Tuple[int, int]
+
+
+@dataclass
+class DynamicReplicaFrameAnnotation:
+    """A dataclass used to load annotations from json."""
+
+    # can be used to join with `SequenceAnnotation`
+    sequence_name: str
+    # 0-based, continuous frame number within sequence
+    frame_number: int
+    # timestamp in seconds from the video start
+    frame_timestamp: float
+
+    image: ImageAnnotation
+    meta: Optional[Dict[str, Any]] = None
+
+    camera_name: Optional[str] = None
+    trajectories: Optional[str] = None
+
+
+class DynamicReplicaDataset(data.Dataset):
+    def __init__(
+        self,
+        root,
+        split="valid",
+        traj_per_sample=256,
+        crop_size=None,
+        sample_len=-1,
+        only_first_n_samples=-1,
+        rgbd_input=False,
+    ):
+        super(DynamicReplicaDataset, self).__init__()
+        self.root = root
+        self.sample_len = sample_len
+        self.split = split
+        self.traj_per_sample = traj_per_sample
+        self.rgbd_input = rgbd_input
+        self.crop_size = crop_size
+        frame_annotations_file = f"frame_annotations_{split}.jgz"
+        self.sample_list = []
+        with gzip.open(
+            os.path.join(root, split, frame_annotations_file), "rt", encoding="utf8"
+        ) as zipfile:
+            frame_annots_list = load_dataclass(
+                zipfile, List[DynamicReplicaFrameAnnotation]
+            )
+        seq_annot = defaultdict(list)
+        for frame_annot in frame_annots_list:
+            if frame_annot.camera_name == "left":
+                seq_annot[frame_annot.sequence_name].append(frame_annot)
+
+        for seq_name in seq_annot.keys():
+            seq_len = len(seq_annot[seq_name])
+
+            step = self.sample_len if self.sample_len > 0 else seq_len
+            counter = 0
+
+            for ref_idx in range(0, seq_len, step):
+                sample = seq_annot[seq_name][ref_idx : ref_idx + step]
+                self.sample_list.append(sample)
+                counter += 1
+                if only_first_n_samples > 0 and counter >= only_first_n_samples:
+                    break
+
+    def __len__(self):
+        return len(self.sample_list)
+
+    def crop(self, rgbs, trajs):
+        T, N, _ = trajs.shape
+
+        S = len(rgbs)
+        H, W = rgbs[0].shape[:2]
+        assert S == T
+
+        H_new = H
+        W_new = W
+
+        # simple random crop
+        y0 = 0 if self.crop_size[0] >= H_new else (H_new - self.crop_size[0]) // 2
+        x0 = 0 if self.crop_size[1] >= W_new else (W_new - self.crop_size[1]) // 2
+        rgbs = [
+            rgb[y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+            for rgb in rgbs
+        ]
+
+        trajs[:, :, 0] -= x0
+        trajs[:, :, 1] -= y0
+
+        return rgbs, trajs
+
+    def __getitem__(self, index):
+        sample = self.sample_list[index]
+        T = len(sample)
+        rgbs, visibilities, traj_2d = [], [], []
+
+        H, W = sample[0].image.size
+        image_size = (H, W)
+
+        for i in range(T):
+            traj_path = os.path.join(
+                self.root, self.split, sample[i].trajectories["path"]
+            )
+            traj = torch.load(traj_path)
+
+            visibilities.append(traj["verts_inds_vis"].numpy())
+
+            rgbs.append(traj["img"].numpy())
+            traj_2d.append(traj["traj_2d"].numpy()[..., :2])
+
+        traj_2d = np.stack(traj_2d)
+        visibility = np.stack(visibilities)
+        T, N, D = traj_2d.shape
+        # subsample trajectories for augmentations
+        visible_inds_sampled = torch.randperm(N)[: self.traj_per_sample]
+
+        traj_2d = traj_2d[:, visible_inds_sampled]
+        visibility = visibility[:, visible_inds_sampled]
+
+        if self.crop_size is not None:
+            rgbs, traj_2d = self.crop(rgbs, traj_2d)
+            H, W, _ = rgbs[0].shape
+            image_size = self.crop_size
+
+        visibility[traj_2d[:, :, 0] > image_size[1] - 1] = False
+        visibility[traj_2d[:, :, 0] < 0] = False
+        visibility[traj_2d[:, :, 1] > image_size[0] - 1] = False
+        visibility[traj_2d[:, :, 1] < 0] = False
+
+        # filter out points that're visible for less than 10 frames
+        visible_inds_resampled = visibility.sum(0) > 10
+        traj_2d = torch.from_numpy(traj_2d[:, visible_inds_resampled])
+        visibility = torch.from_numpy(visibility[:, visible_inds_resampled])
+
+        rgbs = np.stack(rgbs, 0)
+        video = torch.from_numpy(rgbs).reshape(T, H, W, 3).permute(0, 3, 1, 2).float()
+        return CoTrackerData(
+            video=video,
+            trajectory=traj_2d,
+            visibility=visibility,
+            valid=torch.ones(T, N),
+            seq_name=sample[0].sequence_name,
+        )

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/kubric_movif_dataset.py

@@ -0,0 +1,542 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import torch
+import cv2
+
+import imageio
+import numpy as np
+
+from cotracker.datasets.utils import CoTrackerData
+from torchvision.transforms import ColorJitter, GaussianBlur
+from PIL import Image
+from cotracker.models.core.model_utils import smart_cat
+
+
+class CoTrackerDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        data_root,
+        crop_size=(384, 512),
+        seq_len=24,
+        traj_per_sample=768,
+        sample_vis_last_frame=False,
+        use_augs=False,
+    ):
+        super(CoTrackerDataset, self).__init__()
+        np.random.seed(0)
+        torch.manual_seed(0)
+        self.data_root = data_root
+        self.seq_len = seq_len
+        self.traj_per_sample = traj_per_sample
+        self.sample_vis_last_frame = sample_vis_last_frame
+        self.use_augs = use_augs
+        self.crop_size = crop_size
+        # photometric augmentation
+        self.photo_aug = ColorJitter(
+            brightness=0.2, contrast=0.2, saturation=0.2, hue=0.25 / 3.14
+        )
+        self.blur_aug = GaussianBlur(11, sigma=(0.1, 2.0))
+
+        self.blur_aug_prob = 0.25
+        self.color_aug_prob = 0.25
+
+        # occlusion augmentation
+        self.eraser_aug_prob = 0.5
+        self.eraser_bounds = [2, 100]
+        self.eraser_max = 10
+
+        # occlusion augmentation
+        self.replace_aug_prob = 0.5
+        self.replace_bounds = [2, 100]
+        self.replace_max = 10
+
+        # spatial augmentations
+        self.pad_bounds = [0, 100]
+        self.crop_size = crop_size
+        self.resize_lim = [0.25, 2.0]  # sample resizes from here
+        self.resize_delta = 0.2
+        self.max_crop_offset = 50
+
+        self.do_flip = True
+        self.h_flip_prob = 0.5
+        self.v_flip_prob = 0.5
+
+    def getitem_helper(self, index):
+        return NotImplementedError
+
+    def __getitem__(self, index):
+        gotit = False
+
+        sample, gotit = self.getitem_helper(index)
+        if not gotit:
+            print("warning: sampling failed")
+            # fake sample, so we can still collate
+            sample = CoTrackerData(
+                video=torch.zeros(
+                    (self.seq_len, 3, self.crop_size[0], self.crop_size[1])
+                ),
+                trajectory=torch.zeros((self.seq_len, self.traj_per_sample, 2)),
+                visibility=torch.zeros((self.seq_len, self.traj_per_sample)),
+                valid=torch.zeros((self.seq_len, self.traj_per_sample)),
+                # dataset_name="kubric",
+            )
+
+        return sample, gotit
+
+    def add_photometric_augs(self, rgbs, trajs, visibles, eraser=True, replace=True):
+        T, N, _ = trajs.shape
+
+        S = len(rgbs)
+        H, W = rgbs[0].shape[:2]
+        assert S == T
+
+        if eraser:
+            ############ eraser transform (per image after the first) ############
+            rgbs = [rgb.astype(np.float32) for rgb in rgbs]
+            for i in range(1, S):
+                if np.random.rand() < self.eraser_aug_prob:
+                    for _ in range(
+                        np.random.randint(1, self.eraser_max + 1)
+                    ):  # number of times to occlude
+                        xc = np.random.randint(0, W)
+                        yc = np.random.randint(0, H)
+                        dx = np.random.randint(
+                            self.eraser_bounds[0], self.eraser_bounds[1]
+                        )
+                        dy = np.random.randint(
+                            self.eraser_bounds[0], self.eraser_bounds[1]
+                        )
+                        x0 = np.clip(xc - dx / 2, 0, W - 1).round().astype(np.int32)
+                        x1 = np.clip(xc + dx / 2, 0, W - 1).round().astype(np.int32)
+                        y0 = np.clip(yc - dy / 2, 0, H - 1).round().astype(np.int32)
+                        y1 = np.clip(yc + dy / 2, 0, H - 1).round().astype(np.int32)
+
+                        mean_color = np.mean(
+                            rgbs[i][y0:y1, x0:x1, :].reshape(-1, 3), axis=0
+                        )
+                        rgbs[i][y0:y1, x0:x1, :] = mean_color
+
+                        occ_inds = np.logical_and(
+                            np.logical_and(trajs[i, :, 0] >= x0, trajs[i, :, 0] < x1),
+                            np.logical_and(trajs[i, :, 1] >= y0, trajs[i, :, 1] < y1),
+                        )
+                        visibles[i, occ_inds] = 0
+            rgbs = [rgb.astype(np.uint8) for rgb in rgbs]
+
+        if replace:
+            rgbs_alt = [
+                np.array(self.photo_aug(Image.fromarray(rgb)), dtype=np.uint8)
+                for rgb in rgbs
+            ]
+            rgbs_alt = [
+                np.array(self.photo_aug(Image.fromarray(rgb)), dtype=np.uint8)
+                for rgb in rgbs_alt
+            ]
+
+            ############ replace transform (per image after the first) ############
+            rgbs = [rgb.astype(np.float32) for rgb in rgbs]
+            rgbs_alt = [rgb.astype(np.float32) for rgb in rgbs_alt]
+            for i in range(1, S):
+                if np.random.rand() < self.replace_aug_prob:
+                    for _ in range(
+                        np.random.randint(1, self.replace_max + 1)
+                    ):  # number of times to occlude
+                        xc = np.random.randint(0, W)
+                        yc = np.random.randint(0, H)
+                        dx = np.random.randint(
+                            self.replace_bounds[0], self.replace_bounds[1]
+                        )
+                        dy = np.random.randint(
+                            self.replace_bounds[0], self.replace_bounds[1]
+                        )
+                        x0 = np.clip(xc - dx / 2, 0, W - 1).round().astype(np.int32)
+                        x1 = np.clip(xc + dx / 2, 0, W - 1).round().astype(np.int32)
+                        y0 = np.clip(yc - dy / 2, 0, H - 1).round().astype(np.int32)
+                        y1 = np.clip(yc + dy / 2, 0, H - 1).round().astype(np.int32)
+
+                        wid = x1 - x0
+                        hei = y1 - y0
+                        y00 = np.random.randint(0, H - hei)
+                        x00 = np.random.randint(0, W - wid)
+                        fr = np.random.randint(0, S)
+                        rep = rgbs_alt[fr][y00 : y00 + hei, x00 : x00 + wid, :]
+                        rgbs[i][y0:y1, x0:x1, :] = rep
+
+                        occ_inds = np.logical_and(
+                            np.logical_and(trajs[i, :, 0] >= x0, trajs[i, :, 0] < x1),
+                            np.logical_and(trajs[i, :, 1] >= y0, trajs[i, :, 1] < y1),
+                        )
+                        visibles[i, occ_inds] = 0
+            rgbs = [rgb.astype(np.uint8) for rgb in rgbs]
+
+        ############ photometric augmentation ############
+        if np.random.rand() < self.color_aug_prob:
+            # random per-frame amount of aug
+            rgbs = [
+                np.array(self.photo_aug(Image.fromarray(rgb)), dtype=np.uint8)
+                for rgb in rgbs
+            ]
+
+        if np.random.rand() < self.blur_aug_prob:
+            # random per-frame amount of blur
+            rgbs = [
+                np.array(self.blur_aug(Image.fromarray(rgb)), dtype=np.uint8)
+                for rgb in rgbs
+            ]
+
+        return rgbs, trajs, visibles
+
+    def add_spatial_augs(self, rgbs, trajs, visibles, crop_size):
+        T, N, __ = trajs.shape
+
+        S = len(rgbs)
+        H, W = rgbs[0].shape[:2]
+        assert S == T
+
+        rgbs = [rgb.astype(np.float32) for rgb in rgbs]
+
+        ############ spatial transform ############
+
+        # padding
+        pad_x0 = np.random.randint(self.pad_bounds[0], self.pad_bounds[1])
+        pad_x1 = np.random.randint(self.pad_bounds[0], self.pad_bounds[1])
+        pad_y0 = np.random.randint(self.pad_bounds[0], self.pad_bounds[1])
+        pad_y1 = np.random.randint(self.pad_bounds[0], self.pad_bounds[1])
+
+        rgbs = [
+            np.pad(rgb, ((pad_y0, pad_y1), (pad_x0, pad_x1), (0, 0))) for rgb in rgbs
+        ]
+        trajs[:, :, 0] += pad_x0
+        trajs[:, :, 1] += pad_y0
+        H, W = rgbs[0].shape[:2]
+
+        # scaling + stretching
+        scale = np.random.uniform(self.resize_lim[0], self.resize_lim[1])
+        scale_x = scale
+        scale_y = scale
+        H_new = H
+        W_new = W
+
+        scale_delta_x = 0.0
+        scale_delta_y = 0.0
+
+        rgbs_scaled = []
+        for s in range(S):
+            if s == 1:
+                scale_delta_x = np.random.uniform(-self.resize_delta, self.resize_delta)
+                scale_delta_y = np.random.uniform(-self.resize_delta, self.resize_delta)
+            elif s > 1:
+                scale_delta_x = (
+                    scale_delta_x * 0.8
+                    + np.random.uniform(-self.resize_delta, self.resize_delta) * 0.2
+                )
+                scale_delta_y = (
+                    scale_delta_y * 0.8
+                    + np.random.uniform(-self.resize_delta, self.resize_delta) * 0.2
+                )
+            scale_x = scale_x + scale_delta_x
+            scale_y = scale_y + scale_delta_y
+
+            # bring h/w closer
+            scale_xy = (scale_x + scale_y) * 0.5
+            scale_x = scale_x * 0.5 + scale_xy * 0.5
+            scale_y = scale_y * 0.5 + scale_xy * 0.5
+
+            # don't get too crazy
+            scale_x = np.clip(scale_x, 0.2, 2.0)
+            scale_y = np.clip(scale_y, 0.2, 2.0)
+
+            H_new = int(H * scale_y)
+            W_new = int(W * scale_x)
+
+            # make it at least slightly bigger than the crop area,
+            # so that the random cropping can add diversity
+            H_new = np.clip(H_new, crop_size[0] + 10, None)
+            W_new = np.clip(W_new, crop_size[1] + 10, None)
+            # recompute scale in case we clipped
+            scale_x = (W_new - 1) / float(W - 1)
+            scale_y = (H_new - 1) / float(H - 1)
+            rgbs_scaled.append(
+                cv2.resize(rgbs[s], (W_new, H_new), interpolation=cv2.INTER_LINEAR)
+            )
+            trajs[s, :, 0] *= scale_x
+            trajs[s, :, 1] *= scale_y
+        rgbs = rgbs_scaled
+        ok_inds = visibles[0, :] > 0
+        vis_trajs = trajs[:, ok_inds]  # S,?,2
+
+        if vis_trajs.shape[1] > 0:
+            mid_x = np.mean(vis_trajs[0, :, 0])
+            mid_y = np.mean(vis_trajs[0, :, 1])
+        else:
+            mid_y = crop_size[0]
+            mid_x = crop_size[1]
+
+        x0 = int(mid_x - crop_size[1] // 2)
+        y0 = int(mid_y - crop_size[0] // 2)
+
+        offset_x = 0
+        offset_y = 0
+
+        for s in range(S):
+            # on each frame, shift a bit more
+            if s == 1:
+                offset_x = np.random.randint(
+                    -self.max_crop_offset, self.max_crop_offset
+                )
+                offset_y = np.random.randint(
+                    -self.max_crop_offset, self.max_crop_offset
+                )
+            elif s > 1:
+                offset_x = int(
+                    offset_x * 0.8
+                    + np.random.randint(-self.max_crop_offset, self.max_crop_offset + 1)
+                    * 0.2
+                )
+                offset_y = int(
+                    offset_y * 0.8
+                    + np.random.randint(-self.max_crop_offset, self.max_crop_offset + 1)
+                    * 0.2
+                )
+            x0 = x0 + offset_x
+            y0 = y0 + offset_y
+
+            H_new, W_new = rgbs[s].shape[:2]
+            if H_new == crop_size[0]:
+                y0 = 0
+            else:
+                y0 = min(max(0, y0), H_new - crop_size[0] - 1)
+
+            if W_new == crop_size[1]:
+                x0 = 0
+            else:
+                x0 = min(max(0, x0), W_new - crop_size[1] - 1)
+
+            rgbs[s] = rgbs[s][y0 : y0 + crop_size[0], x0 : x0 + crop_size[1]]
+            trajs[s, :, 0] -= x0
+            trajs[s, :, 1] -= y0
+
+        H_new = crop_size[0]
+        W_new = crop_size[1]
+
+        # flip
+        h_flipped = False
+        v_flipped = False
+        if self.do_flip:
+            # h flip
+            if np.random.rand() < self.h_flip_prob:
+                h_flipped = True
+                rgbs = [rgb[:, ::-1] for rgb in rgbs]
+            # v flip
+            if np.random.rand() < self.v_flip_prob:
+                v_flipped = True
+                rgbs = [rgb[::-1] for rgb in rgbs]
+        if h_flipped:
+            trajs[:, :, 0] = W_new - trajs[:, :, 0]
+        if v_flipped:
+            trajs[:, :, 1] = H_new - trajs[:, :, 1]
+        return np.stack(rgbs), trajs
+
+    def crop(self, rgbs, trajs, crop_size):
+        T, N, _ = trajs.shape
+
+        S = len(rgbs)
+        H, W = rgbs[0].shape[:2]
+        assert S == T
+
+        ############ spatial transform ############
+
+        H_new = H
+        W_new = W
+
+        # simple random crop
+        y0 = 0 if crop_size[0] >= H_new else (H_new - crop_size[0]) // 2
+        # np.random.randint(0,
+        x0 = 0 if crop_size[1] >= W_new else np.random.randint(0, W_new - crop_size[1])
+        rgbs = [rgb[y0 : y0 + crop_size[0], x0 : x0 + crop_size[1]] for rgb in rgbs]
+
+        trajs[:, :, 0] -= x0
+        trajs[:, :, 1] -= y0
+
+        return np.stack(rgbs), trajs
+
+
+class KubricMovifDataset(CoTrackerDataset):
+    def __init__(
+        self,
+        data_root,
+        crop_size=(384, 512),
+        seq_len=24,
+        traj_per_sample=768,
+        sample_vis_last_frame=False,
+        use_augs=False,
+        random_seq_len=False,
+        random_frame_rate=False,
+        random_number_traj=False,
+        split="train",
+    ):
+        super(KubricMovifDataset, self).__init__(
+            data_root=data_root,
+            crop_size=crop_size,
+            seq_len=seq_len,
+            traj_per_sample=traj_per_sample,
+            sample_vis_last_frame=sample_vis_last_frame,
+            use_augs=use_augs,
+        )
+        self.random_seq_len = random_seq_len
+        self.random_frame_rate = random_frame_rate
+        self.random_number_traj = random_number_traj
+        self.pad_bounds = [0, 25]
+        self.resize_lim = [0.75, 1.25]  # sample resizes from here
+        self.resize_delta = 0.05
+        self.max_crop_offset = 15
+        self.split = split
+
+        self.seq_names = [
+            fname
+            for fname in os.listdir(data_root)
+            if os.path.isdir(os.path.join(data_root, fname))
+        ]
+        if self.split == "valid":
+            self.seq_names = self.seq_names[:30]
+            assert use_augs == False
+
+        print("found %d unique videos in %s" % (len(self.seq_names), self.data_root))
+
+    def getitem_helper(self, index):
+        gotit = True
+        seq_name = self.seq_names[index]
+        npy_path = os.path.join(self.data_root, seq_name, seq_name + ".npy")
+        rgb_path = os.path.join(self.data_root, seq_name, "frames")
+
+        img_paths = sorted(os.listdir(rgb_path))
+        rgbs = []
+        for i, img_path in enumerate(img_paths):
+            rgbs.append(imageio.v2.imread(os.path.join(rgb_path, img_path)))
+
+        rgbs = np.stack(rgbs)
+        annot_dict = np.load(npy_path, allow_pickle=True).item()
+        traj_2d = annot_dict["coords"]
+        visibility = annot_dict["visibility"]
+
+        frame_rate = 1
+        final_num_traj = self.traj_per_sample
+        crop_size = self.crop_size
+
+        # random crop
+        min_num_traj = 1
+        assert self.traj_per_sample >= min_num_traj
+        if self.random_seq_len and self.random_number_traj:
+            final_num_traj = np.random.randint(min_num_traj, self.traj_per_sample)
+            alpha = final_num_traj / float(self.traj_per_sample)
+            seq_len = int(alpha * 10 + (1 - alpha) * self.seq_len)
+            seq_len = np.random.randint(seq_len - 2, seq_len + 2)
+            if self.random_frame_rate:
+                frame_rate = np.random.randint(1, int((120 / seq_len)) + 1)
+        elif self.random_number_traj:
+            final_num_traj = np.random.randint(min_num_traj, self.traj_per_sample)
+            alpha = final_num_traj / float(self.traj_per_sample)
+            seq_len = 8 * int(alpha * 2 + (1 - alpha) * self.seq_len // 8)
+            # seq_len = np.random.randint(seq_len , seq_len + 2)
+            if self.random_frame_rate:
+                frame_rate = np.random.randint(1, int((120 / seq_len)) + 1)
+        elif self.random_seq_len:
+            seq_len = np.random.randint(int(self.seq_len / 2), self.seq_len)
+            if self.random_frame_rate:
+                frame_rate = np.random.randint(1, int((120 / seq_len)) + 1)
+        else:
+            seq_len = self.seq_len
+            if self.random_frame_rate:
+                frame_rate = np.random.randint(1, int((120 / seq_len)) + 1)
+
+        traj_2d = np.transpose(traj_2d, (1, 0, 2))
+        visibility = np.transpose(np.logical_not(visibility), (1, 0))
+
+        no_augs = False
+        if seq_len < len(rgbs):
+            if seq_len * frame_rate < len(rgbs):
+                start_ind = np.random.choice(len(rgbs) - (seq_len * frame_rate), 1)[0]
+            else:
+                start_ind = 0
+            rgbs = rgbs[start_ind : start_ind + seq_len * frame_rate : frame_rate]
+            traj_2d = traj_2d[start_ind : start_ind + seq_len * frame_rate : frame_rate]
+            visibility = visibility[
+                start_ind : start_ind + seq_len * frame_rate : frame_rate
+            ]
+
+        assert seq_len <= len(rgbs)
+
+        if not no_augs:
+            if self.use_augs:
+                rgbs, traj_2d, visibility = self.add_photometric_augs(
+                    rgbs, traj_2d, visibility, replace=False
+                )
+                rgbs, traj_2d = self.add_spatial_augs(
+                    rgbs, traj_2d, visibility, crop_size
+                )
+            else:
+                rgbs, traj_2d = self.crop(rgbs, traj_2d, crop_size)
+
+        visibility[traj_2d[:, :, 0] > crop_size[1] - 1] = False
+        visibility[traj_2d[:, :, 0] < 0] = False
+        visibility[traj_2d[:, :, 1] > crop_size[0] - 1] = False
+        visibility[traj_2d[:, :, 1] < 0] = False
+
+        visibility = torch.from_numpy(visibility)
+        traj_2d = torch.from_numpy(traj_2d)
+
+        crop_tensor = torch.tensor(crop_size).flip(0)[None, None] / 2.0
+        close_pts_inds = torch.all(
+            torch.linalg.vector_norm(traj_2d[..., :2] - crop_tensor, dim=-1) < 1000.0,
+            dim=0,
+        )
+        traj_2d = traj_2d[:, close_pts_inds]
+        visibility = visibility[:, close_pts_inds]
+
+        visibile_pts_first_frame_inds = (visibility[0]).nonzero(as_tuple=False)[:, 0]
+
+        visibile_pts_mid_frame_inds = (visibility[seq_len // 2]).nonzero(
+            as_tuple=False
+        )[:, 0]
+        visibile_pts_inds = torch.cat(
+            (visibile_pts_first_frame_inds, visibile_pts_mid_frame_inds), dim=0
+        )
+        if self.sample_vis_last_frame:
+            visibile_pts_last_frame_inds = (visibility[seq_len - 1]).nonzero(
+                as_tuple=False
+            )[:, 0]
+            visibile_pts_inds = torch.cat(
+                (visibile_pts_inds, visibile_pts_last_frame_inds), dim=0
+            )
+        point_inds = torch.randperm(len(visibile_pts_inds))[: self.traj_per_sample]
+        if len(point_inds) < self.traj_per_sample:
+            gotit = False
+
+        visible_inds_sampled = visibile_pts_inds[point_inds]
+
+        trajs = traj_2d[:, visible_inds_sampled].float()
+        visibles = visibility[:, visible_inds_sampled]
+        valids = torch.ones_like(visibles)
+
+        trajs = trajs[:, :final_num_traj]
+        visibles = visibles[:, :final_num_traj]
+        valids = valids[:, :final_num_traj]
+
+        rgbs = torch.from_numpy(rgbs).permute(0, 3, 1, 2).float()
+
+        sample = CoTrackerData(
+            video=rgbs,
+            trajectory=trajs,
+            visibility=visibles,
+            valid=valids,
+            seq_name=seq_name,
+        )
+        return sample, gotit
+
+    def __len__(self):
+        return len(self.seq_names)

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/real_dataset.py

@@ -0,0 +1,282 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import torch
+import json
+import cv2
+import math
+import imageio
+import numpy as np
+
+from cotracker.datasets.utils import CoTrackerData
+from torchvision.transforms import ColorJitter, GaussianBlur
+from PIL import Image
+from cotracker.models.core.model_utils import smart_cat
+from torchvision.io import read_video
+import torchvision
+from cotracker.datasets.utils import collate_fn, collate_fn_train, dataclass_to_cuda_
+import torchvision.transforms.functional as F
+
+
+class RealDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        crop_size=(384, 512),
+        seq_len=24,
+        traj_per_sample=768,
+        random_frame_rate=False,
+        random_seq_len=False,
+        data_splits=[0],
+        random_resize=False,
+        limit_samples=10000,
+    ):
+        super(RealDataset, self).__init__()
+        np.random.seed(0)
+        torch.manual_seed(0)
+        raise ValueError(f"This dataset wasn't released. You should collect your own dataset of real videos before training with this dataset class.")
+
+        stopwords = set(
+            [
+                "river",
+                "water",
+                "shore",
+                "lake",
+                "sea",
+                "ocean",
+                "silhouette",
+                "matte",
+                "online",
+                "virtual",
+                "meditation",
+                "artwork",
+                "drawing",
+                "animation",
+                "abstract",
+                "background",
+                "concept",
+                "cartoon",
+                "symbolic",
+                "painting",
+                "sketch",
+                "fireworks",
+                "fire",
+                "sky",
+                "darkness",
+                "timelapse",
+                "time-lapse",
+                "cgi",
+                "computer",
+                "computer-generated",
+                "drawing",
+                "draw",
+                "cgi",
+                "animate",
+                "cartoon",
+                "static",
+                "abstract",
+                "abstraction",
+                "3d",
+                "fandom",
+                "fantasy",
+                "graphics",
+                "cell",
+                "holographic",
+                "generated",
+                "generation" "telephoto",
+                "animated",
+                "disko",
+                "generate" "2d",
+                "3d",
+                "geometric",
+                "geometry",
+                "render",
+                "rendering",
+                "timelapse",
+                "slomo",
+                "slo",
+                "wallpaper",
+                "pattern",
+                "tile",
+                "generated",
+                "chroma",
+                "www",
+                "http",
+                "cannabis",
+                "loop",
+                "cycle",
+                "alpha",
+                "abstract",
+                "concept",
+                "digital",
+                "graphic",
+                "skies",
+                "fountain",
+                "train",
+                "rapid",
+                "fast",
+                "quick",
+                "vfx",
+                "effect",
+            ]
+        )
+
+        def no_stopwords_in_key(key, stopwords):
+            for s in stopwords:
+                if s in key.split(","):
+                    return False
+            return True
+
+        filelist_all = []
+
+        for part in data_splits:
+            filelist = np.load('YOUR FILELIST')
+            captions = np.load('YOUR CAPTIONS')
+            keywords = np.load('YOUR KEYWORDS')
+
+            filtered_seqs_motion = [
+                i
+                for i, key in enumerate(keywords)
+                if "motion" in key.split(",")
+                and (
+                    "man" in key.split(",")
+                    or "woman" in key.split(",")
+                    or "animal" in key.split(",")
+                    or "child" in key.split(",")
+                )
+                and no_stopwords_in_key(key, stopwords)
+            ]
+            print("filtered_seqs_motion", len(filtered_seqs_motion))
+            filtered_seqs = filtered_seqs_motion
+
+            print(f"filtered_seqs {part}", len(filtered_seqs))
+            filelist_all = filelist_all + filelist[filtered_seqs].tolist()
+
+            if len(filelist_all) > limit_samples:
+                break
+
+        self.filelist = filelist_all[:limit_samples]
+        print(f"found {len(self.filelist)} unique videos")
+        self.traj_per_sample = traj_per_sample
+        self.crop_size = crop_size
+        self.seq_len = seq_len
+        self.random_frame_rate = random_frame_rate
+        self.random_resize = random_resize
+        self.random_seq_len = random_seq_len
+
+    def crop(self, rgbs):
+        S = len(rgbs)
+
+        H, W = rgbs.shape[2:]
+
+        H_new = H
+        W_new = W
+
+        # simple random crop
+        y0 = (
+            0
+            if self.crop_size[0] >= H_new
+            else np.random.randint(0, H_new - self.crop_size[0])
+        )
+        x0 = (
+            0
+            if self.crop_size[1] >= W_new
+            else np.random.randint(0, W_new - self.crop_size[1])
+        )
+        rgbs = [
+            rgb[:, y0 : y0 + self.crop_size[0], x0 : x0 + self.crop_size[1]]
+            for rgb in rgbs
+        ]
+
+        return torch.stack(rgbs)
+
+    def __getitem__(self, index):
+        gotit = False
+
+        sample, gotit = self.getitem_helper(index)
+        if not gotit:
+            print("warning: sampling failed")
+            # fake sample, so we can still collate
+            sample = CoTrackerData(
+                video=torch.zeros(
+                    (self.seq_len, 3, self.crop_size[0], self.crop_size[1])
+                ),
+                trajectory=torch.ones(1, 1, 1, 2),
+                visibility=torch.ones(1, 1, 1),
+                valid=torch.ones(1, 1, 1),
+            )
+
+        return sample, gotit
+
+    def sample_h_w(self):
+        area = np.random.uniform(0.6, 1)
+        a1 = np.random.uniform(area, 1)
+        a2 = np.random.uniform(area, 1)
+        h = (a1 + a2) / 2.0
+        w = area / h
+        return h, w
+
+    def getitem_helper(self, index):
+        gotit = True
+        video_path = self.filelist[index]
+
+        rgbs, _, _ = read_video(str(video_path), output_format="TCHW", pts_unit="sec")
+        if rgbs.numel() == 0:
+            return None, False
+        seq_name = video_path
+        frame_rate = 1
+
+        if self.random_seq_len:
+            seq_len = np.random.randint(int(self.seq_len / 2), self.seq_len)
+        else:
+            seq_len = self.seq_len
+
+        while len(rgbs) < seq_len:
+            rgbs = torch.cat([rgbs, rgbs.flip(0)])
+        if seq_len < 8:
+            print("seq_len < 8, return NONE")
+            return None, False
+        if self.random_frame_rate:
+            max_frame_rate = min(4, int((len(rgbs) / seq_len)))
+            if max_frame_rate > 1:
+                frame_rate = np.random.randint(1, max_frame_rate)
+
+        if seq_len * frame_rate < len(rgbs):
+            start_ind = np.random.choice(len(rgbs) - (seq_len * frame_rate), 1)[0]
+        else:
+            start_ind = 0
+        rgbs = rgbs[start_ind : start_ind + seq_len * frame_rate : frame_rate]
+
+        assert seq_len <= len(rgbs)
+
+        if self.random_resize and np.random.rand() < 0.5:
+            video = []
+            rgbs = rgbs.permute(0, 2, 3, 1).numpy()
+
+            for i in range(len(rgbs)):
+                rgb = cv2.resize(
+                    rgbs[i],
+                    (self.crop_size[1], self.crop_size[0]),
+                    interpolation=cv2.INTER_LINEAR,
+                )
+                video.append(rgb)
+            video = torch.tensor(np.stack(video)).permute(0, 3, 1, 2)
+
+        else:
+            video = self.crop(rgbs)
+
+        sample = CoTrackerData(
+            video=video,
+            trajectory=torch.ones(seq_len, self.traj_per_sample, 2),
+            visibility=torch.ones(seq_len, self.traj_per_sample),
+            valid=torch.ones(seq_len, self.traj_per_sample),
+            seq_name=seq_name,
+        )
+
+        return sample, gotit
+
+    def __len__(self):
+        return len(self.filelist)

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/tap_vid_datasets.py

@@ -0,0 +1,244 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import io
+import glob
+import torch
+import pickle
+import numpy as np
+import mediapy as media
+import random
+from PIL import Image
+from typing import Mapping, Tuple, Union
+
+from cotracker.datasets.utils import CoTrackerData
+
+DatasetElement = Mapping[str, Mapping[str, Union[np.ndarray, str]]]
+
+
+def resize_video(video: np.ndarray, output_size: Tuple[int, int]) -> np.ndarray:
+    """Resize a video to output_size."""
+    # If you have a GPU, consider replacing this with a GPU-enabled resize op,
+    # such as a jitted jax.image.resize.  It will make things faster.
+    return media.resize_video(video, output_size)
+
+
+def sample_queries_first(
+    target_occluded: np.ndarray,
+    target_points: np.ndarray,
+    frames: np.ndarray,
+) -> Mapping[str, np.ndarray]:
+    """Package a set of frames and tracks for use in TAPNet evaluations.
+    Given a set of frames and tracks with no query points, use the first
+    visible point in each track as the query.
+    Args:
+      target_occluded: Boolean occlusion flag, of shape [n_tracks, n_frames],
+        where True indicates occluded.
+      target_points: Position, of shape [n_tracks, n_frames, 2], where each point
+        is [x,y] scaled between 0 and 1.
+      frames: Video tensor, of shape [n_frames, height, width, 3].  Scaled between
+        -1 and 1.
+    Returns:
+      A dict with the keys:
+        video: Video tensor of shape [1, n_frames, height, width, 3]
+        query_points: Query points of shape [1, n_queries, 3] where
+          each point is [t, y, x] scaled to the range [-1, 1]
+        target_points: Target points of shape [1, n_queries, n_frames, 2] where
+          each point is [x, y] scaled to the range [-1, 1]
+    """
+    valid = np.sum(~target_occluded, axis=1) > 0
+    target_points = target_points[valid, :]
+    target_occluded = target_occluded[valid, :]
+
+    query_points = []
+    for i in range(target_points.shape[0]):
+        index = np.where(target_occluded[i] == 0)[0][0]
+        x, y = target_points[i, index, 0], target_points[i, index, 1]
+        query_points.append(np.array([index, y, x]))  # [t, y, x]
+    query_points = np.stack(query_points, axis=0)
+
+    return {
+        "video": frames[np.newaxis, ...],
+        "query_points": query_points[np.newaxis, ...],
+        "target_points": target_points[np.newaxis, ...],
+        "occluded": target_occluded[np.newaxis, ...],
+    }
+
+
+def sample_queries_strided(
+    target_occluded: np.ndarray,
+    target_points: np.ndarray,
+    frames: np.ndarray,
+    query_stride: int = 5,
+) -> Mapping[str, np.ndarray]:
+    """Package a set of frames and tracks for use in TAPNet evaluations.
+
+    Given a set of frames and tracks with no query points, sample queries
+    strided every query_stride frames, ignoring points that are not visible
+    at the selected frames.
+
+    Args:
+      target_occluded: Boolean occlusion flag, of shape [n_tracks, n_frames],
+        where True indicates occluded.
+      target_points: Position, of shape [n_tracks, n_frames, 2], where each point
+        is [x,y] scaled between 0 and 1.
+      frames: Video tensor, of shape [n_frames, height, width, 3].  Scaled between
+        -1 and 1.
+      query_stride: When sampling query points, search for un-occluded points
+        every query_stride frames and convert each one into a query.
+
+    Returns:
+      A dict with the keys:
+        video: Video tensor of shape [1, n_frames, height, width, 3].  The video
+          has floats scaled to the range [-1, 1].
+        query_points: Query points of shape [1, n_queries, 3] where
+          each point is [t, y, x] scaled to the range [-1, 1].
+        target_points: Target points of shape [1, n_queries, n_frames, 2] where
+          each point is [x, y] scaled to the range [-1, 1].
+        trackgroup: Index of the original track that each query point was
+          sampled from.  This is useful for visualization.
+    """
+    tracks = []
+    occs = []
+    queries = []
+    trackgroups = []
+    total = 0
+    trackgroup = np.arange(target_occluded.shape[0])
+    for i in range(0, target_occluded.shape[1], query_stride):
+        mask = target_occluded[:, i] == 0
+        query = np.stack(
+            [
+                i * np.ones(target_occluded.shape[0:1]),
+                target_points[:, i, 1],
+                target_points[:, i, 0],
+            ],
+            axis=-1,
+        )
+        queries.append(query[mask])
+        tracks.append(target_points[mask])
+        occs.append(target_occluded[mask])
+        trackgroups.append(trackgroup[mask])
+        total += np.array(np.sum(target_occluded[:, i] == 0))
+
+    return {
+        "video": frames[np.newaxis, ...],
+        "query_points": np.concatenate(queries, axis=0)[np.newaxis, ...],
+        "target_points": np.concatenate(tracks, axis=0)[np.newaxis, ...],
+        "occluded": np.concatenate(occs, axis=0)[np.newaxis, ...],
+        "trackgroup": np.concatenate(trackgroups, axis=0)[np.newaxis, ...],
+    }
+
+
+class TapVidDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        data_root,
+        dataset_type="davis",
+        resize_to=[256, 256],
+        queried_first=True,
+        fast_eval=False,
+    ):
+        local_random = random.Random()
+        local_random.seed(42)
+        self.fast_eval = fast_eval
+        self.dataset_type = dataset_type
+        self.resize_to = resize_to
+        self.queried_first = queried_first
+        if self.dataset_type == "kinetics":
+            all_paths = glob.glob(os.path.join(data_root, "*_of_0010.pkl"))
+            points_dataset = []
+            for pickle_path in all_paths:
+                with open(pickle_path, "rb") as f:
+                    data = pickle.load(f)
+                    points_dataset = points_dataset + data
+            if fast_eval:
+                points_dataset = local_random.sample(points_dataset, 50)
+            self.points_dataset = points_dataset
+
+        elif self.dataset_type == "robotap":
+            all_paths = glob.glob(os.path.join(data_root, "robotap_split*.pkl"))
+            points_dataset = None
+            for pickle_path in all_paths:
+                with open(pickle_path, "rb") as f:
+                    data = pickle.load(f)
+                    if points_dataset is None:
+                        points_dataset = dict(data)
+                    else:
+                        points_dataset.update(data)
+            if fast_eval:
+                points_dataset_keys = local_random.sample(
+                    sorted(points_dataset.keys()), 50
+                )
+                points_dataset = {k: points_dataset[k] for k in points_dataset_keys}
+            self.points_dataset = points_dataset
+            self.video_names = list(self.points_dataset.keys())
+        else:
+            with open(data_root, "rb") as f:
+                self.points_dataset = pickle.load(f)
+            if self.dataset_type == "davis":
+                self.video_names = list(self.points_dataset.keys())
+            elif self.dataset_type == "stacking":
+                # print("self.points_dataset", self.points_dataset)
+                self.video_names = [i for i in range(len(self.points_dataset))]
+        print("found %d unique videos in %s" % (len(self.points_dataset), data_root))
+
+    def __getitem__(self, index):
+        if self.dataset_type == "davis" or self.dataset_type == "robotap":
+            video_name = self.video_names[index]
+        else:
+            video_name = index
+        video = self.points_dataset[video_name]
+        frames = video["video"]
+
+        if self.fast_eval and frames.shape[0] > 300:
+            return self.__getitem__((index + 1) % self.__len__())
+        if isinstance(frames[0], bytes):
+            # TAP-Vid is stored and JPEG bytes rather than `np.ndarray`s.
+            def decode(frame):
+                byteio = io.BytesIO(frame)
+                img = Image.open(byteio)
+                return np.array(img)
+
+            frames = np.array([decode(frame) for frame in frames])
+
+        target_points = self.points_dataset[video_name]["points"]
+        if self.resize_to is not None:
+            frames = resize_video(frames, self.resize_to)
+            target_points *= np.array(
+                [self.resize_to[1] - 1, self.resize_to[0] - 1]
+            )  # 1 should be mapped to resize_to-1
+        else:
+            target_points *= np.array([frames.shape[2] - 1, frames.shape[1] - 1])
+
+        target_occ = self.points_dataset[video_name]["occluded"]
+        if self.queried_first:
+            converted = sample_queries_first(target_occ, target_points, frames)
+        else:
+            converted = sample_queries_strided(target_occ, target_points, frames)
+        assert converted["target_points"].shape[1] == converted["query_points"].shape[1]
+
+        trajs = (
+            torch.from_numpy(converted["target_points"])[0].permute(1, 0, 2).float()
+        )  # T, N, D
+
+        rgbs = torch.from_numpy(frames).permute(0, 3, 1, 2).float()
+        visibles = torch.logical_not(torch.from_numpy(converted["occluded"]))[
+            0
+        ].permute(
+            1, 0
+        )  # T, N
+        query_points = torch.from_numpy(converted["query_points"])[0]  # T, N
+        return CoTrackerData(
+            rgbs,
+            trajs,
+            visibles,
+            seq_name=str(video_name),
+            query_points=query_points,
+        )
+
+    def __len__(self):
+        return len(self.points_dataset)

torch_hub/facebookresearch_co-tracker_main/cotracker/datasets/utils.py

@@ -0,0 +1,120 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import torch
+import dataclasses
+import torch.nn.functional as F
+from dataclasses import dataclass
+from typing import Any, Optional, Dict
+
+
+@dataclass(eq=False)
+class CoTrackerData:
+    """
+    Dataclass for storing video tracks data.
+    """
+
+    video: torch.Tensor  # B, S, C, H, W
+    trajectory: torch.Tensor  # B, S, N, 2
+    visibility: torch.Tensor  # B, S, N
+    # optional data
+    valid: Optional[torch.Tensor] = None  # B, S, N
+    segmentation: Optional[torch.Tensor] = None  # B, S, 1, H, W
+    seq_name: Optional[str] = None
+    query_points: Optional[torch.Tensor] = None  # TapVID evaluation format
+    transforms: Optional[Dict[str, Any]] = None
+    aug_video: Optional[torch.Tensor] = None
+
+
+def collate_fn(batch):
+    """
+    Collate function for video tracks data.
+    """
+    video = torch.stack([b.video for b in batch], dim=0)
+    trajectory = torch.stack([b.trajectory for b in batch], dim=0)
+    visibility = torch.stack([b.visibility for b in batch], dim=0)
+    query_points = segmentation = None
+    if batch[0].query_points is not None:
+        query_points = torch.stack([b.query_points for b in batch], dim=0)
+    if batch[0].segmentation is not None:
+        segmentation = torch.stack([b.segmentation for b in batch], dim=0)
+    seq_name = [b.seq_name for b in batch]
+
+    return CoTrackerData(
+        video=video,
+        trajectory=trajectory,
+        visibility=visibility,
+        segmentation=segmentation,
+        seq_name=seq_name,
+        query_points=query_points,
+    )
+
+
+def collate_fn_train(batch):
+    """
+    Collate function for video tracks data during training.
+    """
+    gotit = [gotit for _, gotit in batch]
+    video = torch.stack([b.video for b, _ in batch], dim=0)
+    trajectory = torch.stack([b.trajectory for b, _ in batch], dim=0)
+    visibility = torch.stack([b.visibility for b, _ in batch], dim=0)
+    valid = torch.stack([b.valid for b, _ in batch], dim=0)
+    seq_name = [b.seq_name for b, _ in batch]
+    query_points = transforms = aug_video = None
+    if batch[0][0].query_points is not None:
+        query_points = torch.stack([b.query_points for b, _ in batch], dim=0)
+
+    if batch[0][0].transforms is not None:
+        transforms = [b.transforms for b, _ in batch]
+
+    if batch[0][0].aug_video is not None:
+        aug_video = torch.stack([b.aug_video for b, _ in batch], dim=0)
+    return (
+        CoTrackerData(
+            video=video,
+            trajectory=trajectory,
+            visibility=visibility,
+            valid=valid,
+            seq_name=seq_name,
+            query_points=query_points,
+            aug_video=aug_video,
+            transforms=transforms,
+        ),
+        gotit,
+    )
+
+
+def try_to_cuda(t: Any) -> Any:
+    """
+    Try to move the input variable `t` to a cuda device.
+
+    Args:
+        t: Input.
+
+    Returns:
+        t_cuda: `t` moved to a cuda device, if supported.
+    """
+    try:
+        t = t.float().cuda()
+    except AttributeError:
+        pass
+    return t
+
+
+def dataclass_to_cuda_(obj):
+    """
+    Move all contents of a dataclass to cuda inplace if supported.
+
+    Args:
+        batch: Input dataclass.
+
+    Returns:
+        batch_cuda: `batch` moved to a cuda device, if supported.
+    """
+    for f in dataclasses.fields(obj):
+        setattr(obj, f.name, try_to_cuda(getattr(obj, f.name)))
+    return obj

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_dynamic_replica.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: dynamic_replica
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_davis_first.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_davis_first
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_davis_strided.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_davis_strided
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_kinetics_first.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_kinetics_first
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_robotap_first.yaml

@@ -0,0 +1,4 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_robotap_first

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_stacking_first.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_stacking_first
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/configs/eval_tapvid_stacking_strided.yaml

@@ -0,0 +1,6 @@
+defaults:
+  - default_config_eval
+exp_dir: ./outputs/cotracker
+dataset_name: tapvid_stacking_strided
+
+   

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/core/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/core/eval_utils.py

@@ -0,0 +1,138 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+
+from typing import Iterable, Mapping, Tuple, Union
+
+
+def compute_tapvid_metrics(
+    query_points: np.ndarray,
+    gt_occluded: np.ndarray,
+    gt_tracks: np.ndarray,
+    pred_occluded: np.ndarray,
+    pred_tracks: np.ndarray,
+    query_mode: str,
+) -> Mapping[str, np.ndarray]:
+    """Computes TAP-Vid metrics (Jaccard, Pts. Within Thresh, Occ. Acc.)
+    See the TAP-Vid paper for details on the metric computation.  All inputs are
+    given in raster coordinates.  The first three arguments should be the direct
+    outputs of the reader: the 'query_points', 'occluded', and 'target_points'.
+    The paper metrics assume these are scaled relative to 256x256 images.
+    pred_occluded and pred_tracks are your algorithm's predictions.
+    This function takes a batch of inputs, and computes metrics separately for
+    each video.  The metrics for the full benchmark are a simple mean of the
+    metrics across the full set of videos.  These numbers are between 0 and 1,
+    but the paper multiplies them by 100 to ease reading.
+    Args:
+       query_points: The query points, an in the format [t, y, x].  Its size is
+         [b, n, 3], where b is the batch size and n is the number of queries
+       gt_occluded: A boolean array of shape [b, n, t], where t is the number
+         of frames.  True indicates that the point is occluded.
+       gt_tracks: The target points, of shape [b, n, t, 2].  Each point is
+         in the format [x, y]
+       pred_occluded: A boolean array of predicted occlusions, in the same
+         format as gt_occluded.
+       pred_tracks: An array of track predictions from your algorithm, in the
+         same format as gt_tracks.
+       query_mode: Either 'first' or 'strided', depending on how queries are
+         sampled.  If 'first', we assume the prior knowledge that all points
+         before the query point are occluded, and these are removed from the
+         evaluation.
+    Returns:
+        A dict with the following keys:
+        occlusion_accuracy: Accuracy at predicting occlusion.
+        pts_within_{x} for x in [1, 2, 4, 8, 16]: Fraction of points
+          predicted to be within the given pixel threshold, ignoring occlusion
+          prediction.
+        jaccard_{x} for x in [1, 2, 4, 8, 16]: Jaccard metric for the given
+          threshold
+        average_pts_within_thresh: average across pts_within_{x}
+        average_jaccard: average across jaccard_{x}
+    """
+
+    metrics = {}
+    # Fixed bug is described in:
+    # https://github.com/facebookresearch/co-tracker/issues/20
+    eye = np.eye(gt_tracks.shape[2], dtype=np.int32)
+
+    if query_mode == "first":
+        # evaluate frames after the query frame
+        query_frame_to_eval_frames = np.cumsum(eye, axis=1) - eye
+    elif query_mode == "strided":
+        # evaluate all frames except the query frame
+        query_frame_to_eval_frames = 1 - eye
+    else:
+        raise ValueError("Unknown query mode " + query_mode)
+
+    query_frame = query_points[..., 0]
+    query_frame = np.round(query_frame).astype(np.int32)
+    evaluation_points = query_frame_to_eval_frames[query_frame] > 0
+
+    # Occlusion accuracy is simply how often the predicted occlusion equals the
+    # ground truth.
+    occ_acc = np.sum(
+        np.equal(pred_occluded, gt_occluded) & evaluation_points,
+        axis=(1, 2),
+    ) / np.sum(evaluation_points)
+    metrics["occlusion_accuracy"] = occ_acc
+
+    # Next, convert the predictions and ground truth positions into pixel
+    # coordinates.
+    visible = np.logical_not(gt_occluded)
+    pred_visible = np.logical_not(pred_occluded)
+    all_frac_within = []
+    all_jaccard = []
+    for thresh in [1, 2, 4, 8, 16]:
+        # True positives are points that are within the threshold and where both
+        # the prediction and the ground truth are listed as visible.
+        within_dist = np.sum(
+            np.square(pred_tracks - gt_tracks),
+            axis=-1,
+        ) < np.square(thresh)
+        is_correct = np.logical_and(within_dist, visible)
+
+        # Compute the frac_within_threshold, which is the fraction of points
+        # within the threshold among points that are visible in the ground truth,
+        # ignoring whether they're predicted to be visible.
+        count_correct = np.sum(
+            is_correct & evaluation_points,
+            axis=(1, 2),
+        )
+        count_visible_points = np.sum(visible & evaluation_points, axis=(1, 2))
+        frac_correct = count_correct / count_visible_points
+        metrics["pts_within_" + str(thresh)] = frac_correct
+        all_frac_within.append(frac_correct)
+
+        true_positives = np.sum(
+            is_correct & pred_visible & evaluation_points, axis=(1, 2)
+        )
+
+        # The denominator of the jaccard metric is the true positives plus
+        # false positives plus false negatives.  However, note that true positives
+        # plus false negatives is simply the number of points in the ground truth
+        # which is easier to compute than trying to compute all three quantities.
+        # Thus we just add the number of points in the ground truth to the number
+        # of false positives.
+        #
+        # False positives are simply points that are predicted to be visible,
+        # but the ground truth is not visible or too far from the prediction.
+        gt_positives = np.sum(visible & evaluation_points, axis=(1, 2))
+        false_positives = (~visible) & pred_visible
+        false_positives = false_positives | ((~within_dist) & pred_visible)
+        false_positives = np.sum(false_positives & evaluation_points, axis=(1, 2))
+        jaccard = true_positives / (gt_positives + false_positives)
+        metrics["jaccard_" + str(thresh)] = jaccard
+        all_jaccard.append(jaccard)
+    metrics["average_jaccard"] = np.mean(
+        np.stack(all_jaccard, axis=1),
+        axis=1,
+    )
+    metrics["average_pts_within_thresh"] = np.mean(
+        np.stack(all_frac_within, axis=1),
+        axis=1,
+    )
+    return metrics

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/core/evaluator.py

@@ -0,0 +1,288 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from collections import defaultdict
+import os
+from typing import Optional
+import torch
+from tqdm import tqdm
+import numpy as np
+
+from torch.utils.tensorboard import SummaryWriter
+from cotracker.datasets.utils import dataclass_to_cuda_
+from cotracker.utils.visualizer import Visualizer
+from cotracker.models.core.model_utils import reduce_masked_mean
+from cotracker.evaluation.core.eval_utils import compute_tapvid_metrics
+from cotracker.predictor import CoTrackerOnlinePredictor
+from cotracker.models.core.cotracker.cotracker3_offline import CoTrackerThreeOffline
+from cotracker.models.core.cotracker.cotracker3_online import CoTrackerThreeOnline
+import logging
+
+
+class Evaluator:
+    """
+    A class defining the CoTracker evaluator.
+    """
+
+    def __init__(self, exp_dir) -> None:
+        # Visualization
+        self.exp_dir = exp_dir
+        os.makedirs(exp_dir, exist_ok=True)
+        self.visualization_filepaths = defaultdict(lambda: defaultdict(list))
+        self.visualize_dir = os.path.join(exp_dir, "visualisations")
+
+    def compute_metrics(self, metrics, sample, pred_trajectory, dataset_name):
+        if isinstance(pred_trajectory, tuple):
+            pred_trajectory, pred_visibility = pred_trajectory
+        else:
+            pred_visibility = None
+        if "tapvid" in dataset_name:
+            B, T, N, D = sample.trajectory.shape
+            traj = sample.trajectory.clone()
+            thr = 0.6
+
+            if pred_visibility is None:
+                logging.warning("visibility is NONE")
+                pred_visibility = torch.zeros_like(sample.visibility)
+
+            if not pred_visibility.dtype == torch.bool:
+                pred_visibility = pred_visibility > thr
+
+            query_points = sample.query_points.clone().cpu().numpy()
+
+            pred_visibility = pred_visibility[:, :, :N]
+            pred_trajectory = pred_trajectory[:, :, :N]
+
+            gt_tracks = traj.permute(0, 2, 1, 3).cpu().numpy()
+            gt_occluded = (
+                torch.logical_not(sample.visibility.clone().permute(0, 2, 1))
+                .cpu()
+                .numpy()
+            )
+
+            pred_occluded = (
+                torch.logical_not(pred_visibility.clone().permute(0, 2, 1))
+                .cpu()
+                .numpy()
+            )
+            pred_tracks = pred_trajectory.permute(0, 2, 1, 3).cpu().numpy()
+
+            out_metrics = compute_tapvid_metrics(
+                query_points,
+                gt_occluded,
+                gt_tracks,
+                pred_occluded,
+                pred_tracks,
+                query_mode="strided" if "strided" in dataset_name else "first",
+            )
+
+            metrics[sample.seq_name[0]] = out_metrics
+            for metric_name in out_metrics.keys():
+                if "avg" not in metrics:
+                    metrics["avg"] = {}
+                metrics["avg"][metric_name] = np.mean(
+                    [v[metric_name] for k, v in metrics.items() if k != "avg"]
+                )
+
+            logging.info(f"Metrics: {out_metrics}")
+            logging.info(f"avg: {metrics['avg']}")
+            print("metrics", out_metrics)
+            print("avg", metrics["avg"])
+        elif dataset_name == "dynamic_replica" or dataset_name == "pointodyssey":
+            *_, N, _ = sample.trajectory.shape
+            B, T, N = sample.visibility.shape
+            H, W = sample.video.shape[-2:]
+            device = sample.video.device
+
+            out_metrics = {}
+
+            d_vis_sum = d_occ_sum = d_sum_all = 0.0
+            thrs = [1, 2, 4, 8, 16]
+            sx_ = (W - 1) / 255.0
+            sy_ = (H - 1) / 255.0
+            sc_py = np.array([sx_, sy_]).reshape([1, 1, 2])
+            sc_pt = torch.from_numpy(sc_py).float().to(device)
+            __, first_visible_inds = torch.max(sample.visibility, dim=1)
+
+            frame_ids_tensor = torch.arange(T, device=device)[None, :, None].repeat(
+                B, 1, N
+            )
+            start_tracking_mask = frame_ids_tensor > (first_visible_inds.unsqueeze(1))
+
+            for thr in thrs:
+                d_ = (
+                    torch.norm(
+                        pred_trajectory[..., :2] / sc_pt
+                        - sample.trajectory[..., :2] / sc_pt,
+                        dim=-1,
+                    )
+                    < thr
+                ).float()  # B,S-1,N
+                d_occ = (
+                    reduce_masked_mean(
+                        d_, (1 - sample.visibility) * start_tracking_mask
+                    ).item()
+                    * 100.0
+                )
+                d_occ_sum += d_occ
+                out_metrics[f"accuracy_occ_{thr}"] = d_occ
+
+                d_vis = (
+                    reduce_masked_mean(
+                        d_, sample.visibility * start_tracking_mask
+                    ).item()
+                    * 100.0
+                )
+                d_vis_sum += d_vis
+                out_metrics[f"accuracy_vis_{thr}"] = d_vis
+
+                d_all = reduce_masked_mean(d_, start_tracking_mask).item() * 100.0
+                d_sum_all += d_all
+                out_metrics[f"accuracy_{thr}"] = d_all
+
+            d_occ_avg = d_occ_sum / len(thrs)
+            d_vis_avg = d_vis_sum / len(thrs)
+            d_all_avg = d_sum_all / len(thrs)
+
+            sur_thr = 50
+            dists = torch.norm(
+                pred_trajectory[..., :2] / sc_pt - sample.trajectory[..., :2] / sc_pt,
+                dim=-1,
+            )  # B,S,N
+            dist_ok = 1 - (dists > sur_thr).float() * sample.visibility  # B,S,N
+            survival = torch.cumprod(dist_ok, dim=1)  # B,S,N
+            out_metrics["survival"] = torch.mean(survival).item() * 100.0
+
+            out_metrics["accuracy_occ"] = d_occ_avg
+            out_metrics["accuracy_vis"] = d_vis_avg
+            out_metrics["accuracy"] = d_all_avg
+
+            metrics[sample.seq_name[0]] = out_metrics
+            for metric_name in out_metrics.keys():
+                if "avg" not in metrics:
+                    metrics["avg"] = {}
+                metrics["avg"][metric_name] = float(
+                    np.mean([v[metric_name] for k, v in metrics.items() if k != "avg"])
+                )
+
+            logging.info(f"Metrics: {out_metrics}")
+            logging.info(f"avg: {metrics['avg']}")
+            print("metrics", out_metrics)
+            print("avg", metrics["avg"])
+
+    @torch.no_grad()
+    def evaluate_sequence(
+        self,
+        model,
+        test_dataloader: torch.utils.data.DataLoader,
+        dataset_name: str,
+        train_mode=False,
+        visualize_every: int = 50,
+        writer: Optional[SummaryWriter] = None,
+        step: Optional[int] = 0,
+    ):
+        metrics = {}
+
+        vis = Visualizer(
+            save_dir=self.exp_dir,
+            fps=7,
+        )
+
+        for ind, sample in enumerate(tqdm(test_dataloader)):
+            if isinstance(sample, tuple):
+                sample, gotit = sample
+                if not all(gotit):
+                    print("batch is None")
+                    continue
+            if torch.cuda.is_available():
+                dataclass_to_cuda_(sample)
+                device = torch.device("cuda")
+            else:
+                device = torch.device("cpu")
+
+            if (
+                not train_mode
+                and hasattr(model, "sequence_len")
+                and (sample.visibility[:, : model.sequence_len].sum() == 0)
+            ):
+                print(f"skipping batch {ind}")
+                continue
+
+            if "tapvid" in dataset_name:
+                queries = sample.query_points.clone().float()
+
+                queries = torch.stack(
+                    [
+                        queries[:, :, 0],
+                        queries[:, :, 2],
+                        queries[:, :, 1],
+                    ],
+                    dim=2,
+                ).to(device)
+            else:
+                queries = torch.cat(
+                    [
+                        torch.zeros_like(sample.trajectory[:, 0, :, :1]),
+                        sample.trajectory[:, 0],
+                    ],
+                    dim=2,
+                ).to(device)
+
+            if isinstance(model.model, CoTrackerThreeOnline):
+                online_model = CoTrackerOnlinePredictor(checkpoint=None)
+                online_model.model = model.model
+                online_model.step = model.model.window_len // 2
+                online_model(
+                    video_chunk=sample.video,
+                    is_first_step=True,
+                    queries=queries,
+                    add_support_grid=False,
+                )
+                # Process the video
+                for ind in range(
+                    0, sample.video.shape[1] - online_model.step, online_model.step
+                ):
+                    pred_tracks, pred_visibility = online_model(
+                        video_chunk=sample.video[:, ind : ind + online_model.step * 2],
+                        add_support_grid=False,
+                        grid_size=0,
+                    )  # B T N 2,  B T N 1
+                pred_tracks = (pred_tracks, pred_visibility)
+            else:
+                pred_tracks = model(sample.video, queries)
+
+            if "strided" in dataset_name:
+                inv_video = sample.video.flip(1).clone()
+                inv_queries = queries.clone()
+                inv_queries[:, :, 0] = inv_video.shape[1] - inv_queries[:, :, 0] - 1
+
+                pred_trj, pred_vsb = pred_tracks
+                inv_pred_trj, inv_pred_vsb = model(inv_video, inv_queries)
+
+                inv_pred_trj = inv_pred_trj.flip(1)
+                inv_pred_vsb = inv_pred_vsb.flip(1)
+
+                mask = pred_trj == 0
+
+                pred_trj[mask] = inv_pred_trj[mask]
+                pred_vsb[mask[:, :, :, 0]] = inv_pred_vsb[mask[:, :, :, 0]]
+
+                pred_tracks = pred_trj, pred_vsb
+
+            if dataset_name == "badja" or dataset_name == "fastcapture":
+                seq_name = sample.seq_name[0]
+            else:
+                seq_name = str(ind)
+            if ind % visualize_every == 0:
+                vis.visualize(
+                    sample.video,
+                    pred_tracks[0] if isinstance(pred_tracks, tuple) else pred_tracks,
+                    filename=dataset_name + "_" + seq_name,
+                    writer=writer,
+                    step=step,
+                )
+            self.compute_metrics(metrics, sample, pred_tracks, dataset_name)
+        return metrics

torch_hub/facebookresearch_co-tracker_main/cotracker/evaluation/evaluate.py

@@ -0,0 +1,190 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import json
+import os
+import hydra
+import numpy as np
+import torch
+
+from typing import Optional
+from dataclasses import dataclass, field
+
+from omegaconf import OmegaConf
+
+from cotracker.datasets.utils import collate_fn
+from cotracker.models.evaluation_predictor import EvaluationPredictor
+
+from cotracker.evaluation.core.evaluator import Evaluator
+from cotracker.models.build_cotracker import build_cotracker
+
+
+@dataclass(eq=False)
+class DefaultConfig:
+    # Directory where all outputs of the experiment will be saved.
+    exp_dir: str = "./outputs"
+
+    # Name of the dataset to be used for the evaluation.
+    dataset_name: str = "tapvid_davis_first"
+    # The root directory of the dataset.
+    dataset_root: str = "./"
+
+    # Path to the pre-trained model checkpoint to be used for the evaluation.
+    # The default value is the path to a specific CoTracker model checkpoint.
+    checkpoint: str = "./checkpoints/scaled_online.pth"
+    # EvaluationPredictor parameters
+    # The size (N) of the support grid used in the predictor.
+    # The total number of points is (N*N).
+    grid_size: int = 5
+    # The size (N) of the local support grid.
+    local_grid_size: int = 8
+    num_uniformly_sampled_pts: int = 0
+    sift_size: int = 0
+    # A flag indicating whether to evaluate one ground truth point at a time.
+    single_point: bool = False
+    offline_model: bool = False
+    window_len: int = 16
+    # The number of iterative updates for each sliding window.
+    n_iters: int = 6
+
+    seed: int = 0
+    gpu_idx: int = 0
+    local_extent: int = 50
+
+    v2: bool = False
+
+    # Override hydra's working directory to current working dir,
+    # also disable storing the .hydra logs:
+    hydra: dict = field(
+        default_factory=lambda: {
+            "run": {"dir": "."},
+            "output_subdir": None,
+        }
+    )
+
+
+def run_eval(cfg: DefaultConfig):
+    """
+    The function evaluates CoTracker on a specified benchmark dataset based on a provided configuration.
+
+    Args:
+        cfg (DefaultConfig): An instance of DefaultConfig class which includes:
+            - exp_dir (str): The directory path for the experiment.
+            - dataset_name (str): The name of the dataset to be used.
+            - dataset_root (str): The root directory of the dataset.
+            - checkpoint (str): The path to the CoTracker model's checkpoint.
+            - single_point (bool): A flag indicating whether to evaluate one ground truth point at a time.
+            - n_iters (int): The number of iterative updates for each sliding window.
+            - seed (int): The seed for setting the random state for reproducibility.
+            - gpu_idx (int): The index of the GPU to be used.
+    """
+    # Creating the experiment directory if it doesn't exist
+    os.makedirs(cfg.exp_dir, exist_ok=True)
+
+    # Saving the experiment configuration to a .yaml file in the experiment directory
+    cfg_file = os.path.join(cfg.exp_dir, "expconfig.yaml")
+    with open(cfg_file, "w") as f:
+        OmegaConf.save(config=cfg, f=f)
+
+    evaluator = Evaluator(cfg.exp_dir)
+    cotracker_model = build_cotracker(
+        cfg.checkpoint, offline=cfg.offline_model, window_len=cfg.window_len, v2=cfg.v2
+    )
+
+    # Creating the EvaluationPredictor object
+    predictor = EvaluationPredictor(
+        cotracker_model,
+        grid_size=cfg.grid_size,
+        local_grid_size=cfg.local_grid_size,
+        sift_size=cfg.sift_size,
+        single_point=cfg.single_point,
+        num_uniformly_sampled_pts=cfg.num_uniformly_sampled_pts,
+        n_iters=cfg.n_iters,
+        local_extent=cfg.local_extent,
+        interp_shape=(384, 512),
+    )
+
+    if torch.cuda.is_available():
+        predictor.model = predictor.model.cuda()
+
+    # Setting the random seeds
+    torch.manual_seed(cfg.seed)
+    np.random.seed(cfg.seed)
+
+    # Constructing the specified dataset
+    curr_collate_fn = collate_fn
+    if "tapvid" in cfg.dataset_name:
+        from cotracker.datasets.tap_vid_datasets import TapVidDataset
+
+        dataset_type = cfg.dataset_name.split("_")[1]
+        if dataset_type == "davis":
+            data_root = os.path.join(
+                cfg.dataset_root, "tapvid_davis", "tapvid_davis.pkl"
+            )
+        elif dataset_type == "kinetics":
+            data_root = os.path.join(cfg.dataset_root, "tapvid_kinetics")
+        elif dataset_type == "robotap":
+            data_root = os.path.join(cfg.dataset_root, "tapvid_robotap")
+        elif dataset_type == "stacking":
+            data_root = os.path.join(
+                cfg.dataset_root, "tapvid_rgb_stacking", "tapvid_rgb_stacking.pkl"
+            )
+
+        test_dataset = TapVidDataset(
+            dataset_type=dataset_type,
+            data_root=data_root,
+            queried_first=not "strided" in cfg.dataset_name,
+            # resize_to=None,
+        )
+    elif cfg.dataset_name == "dynamic_replica":
+        from cotracker.datasets.dr_dataset import DynamicReplicaDataset
+
+        test_dataset = DynamicReplicaDataset(
+            cfg.dataset_root, sample_len=300, only_first_n_samples=1
+        )
+
+    # Creating the DataLoader object
+    test_dataloader = torch.utils.data.DataLoader(
+        test_dataset,
+        batch_size=1,
+        shuffle=False,
+        num_workers=1,
+        collate_fn=curr_collate_fn,
+    )
+
+    # Timing and conducting the evaluation
+    import time
+
+    start = time.time()
+    evaluate_result = evaluator.evaluate_sequence(
+        predictor, test_dataloader, dataset_name=cfg.dataset_name
+    )
+    end = time.time()
+    print(end - start)
+
+    # Saving the evaluation results to a .json file
+    evaluate_result = evaluate_result["avg"]
+    print("evaluate_result", evaluate_result)
+    result_file = os.path.join(cfg.exp_dir, f"result_eval_.json")
+    evaluate_result["time"] = end - start
+    print(f"Dumping eval results to {result_file}.")
+    with open(result_file, "w") as f:
+        json.dump(evaluate_result, f)
+
+
+cs = hydra.core.config_store.ConfigStore.instance()
+cs.store(name="default_config_eval", node=DefaultConfig)
+
+
+@hydra.main(config_path="./configs/", config_name="default_config_eval")
+def evaluate(cfg: DefaultConfig) -> None:
+    os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID"
+    os.environ["CUDA_VISIBLE_DEVICES"] = str(cfg.gpu_idx)
+    run_eval(cfg)
+
+
+if __name__ == "__main__":
+    evaluate()

torch_hub/facebookresearch_co-tracker_main/cotracker/models/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/models/bootstap_predictor.py

@@ -0,0 +1,65 @@
+import torch
+import torch.nn.functional as F
+
+import sys
+
+import matplotlib.pyplot as plt
+import mediapy as media
+import numpy as np
+from tapnet.torch.tapir_model import TAPIR
+
+
+def postprocess_occlusions(occlusions, expected_dist):
+    visibles = (1 - F.sigmoid(occlusions)) * (1 - F.sigmoid(expected_dist)) > 0.5
+    return visibles
+
+
+class TAPIRPredictor(torch.nn.Module):
+    def __init__(self, bootstap=False, model=None):
+        super().__init__()
+        self.interp_shape = (256, 256)
+        if model is None:
+            if bootstap:
+                checkpoint = "./tapnet/bootstapir_checkpoint.pt"
+                model = TAPIR(pyramid_level=1, extra_convs=True)
+            else:
+                checkpoint = "./tapnet/tapir_checkpoint_panning.pt"
+                model = TAPIR(pyramid_level=0, extra_convs=False)
+            model.load_state_dict(torch.load(checkpoint))
+        self.model = model.eval().to("cuda")
+
+    def forward(self, rgbs, queries=None, grid_size=0, iters=6, eval_depth=False):
+        B, T, C, H, W = rgbs.shape
+        rgbs_ = rgbs.reshape(B * T, C, H, W)
+        rgbs_ = F.interpolate(rgbs_, tuple(self.interp_shape), mode="bilinear")
+        rgbs_ = rgbs_.reshape(B, T, 3, self.interp_shape[0], self.interp_shape[1])
+        rgbs_ = rgbs_[0].permute(0, 2, 3, 1)
+        rgbs_ = (rgbs_ / 255.0) * 2 - 1
+
+        if queries is not None:
+            queries = queries.clone().float()
+            B, N, D = queries.shape
+            assert D == 3
+            assert B == 1
+            queries[:, :, 1] *= self.interp_shape[1] / W
+            queries[:, :, 2] *= self.interp_shape[0] / H
+            queries = torch.stack(
+                [queries[..., 0], queries[..., 2], queries[..., 1]], dim=-1
+            )
+
+        outputs = self.model(video=rgbs_[None], query_points=queries)
+        tracks, occlusions, expected_dist = (
+            outputs["tracks"],
+            outputs["occlusion"][0],
+            outputs["expected_dist"][0],
+        )
+        visibility = postprocess_occlusions(occlusions, expected_dist)[None].permute(
+            0, 2, 1
+        )
+
+        tracks = tracks.permute(0, 2, 1, 3)
+
+        tracks[:, :, :, 0] *= W / float(self.interp_shape[1])
+        tracks[:, :, :, 1] *= H / float(self.interp_shape[0])
+
+        return tracks, visibility

torch_hub/facebookresearch_co-tracker_main/cotracker/models/build_cotracker.py

@@ -0,0 +1,45 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from cotracker.models.core.cotracker.cotracker import CoTracker2
+from cotracker.models.core.cotracker.cotracker3_offline import CoTrackerThreeOffline
+from cotracker.models.core.cotracker.cotracker3_online import CoTrackerThreeOnline
+
+
+def build_cotracker(
+    checkpoint: str,
+):
+    if checkpoint is None:
+        return build_cotracker()
+    model_name = checkpoint.split("/")[-1].split(".")[0]
+    if model_name == "cotracker":
+        return build_cotracker(checkpoint=checkpoint)
+    else:
+        raise ValueError(f"Unknown model name {model_name}")
+
+
+def build_cotracker(checkpoint=None, offline=True, window_len=16, v2=False):
+    if v2:
+        cotracker = CoTracker2(stride=4, window_len=window_len)
+    else:
+        if offline:
+            cotracker = CoTrackerThreeOffline(
+                stride=4, corr_radius=3, window_len=window_len
+            )
+        else:
+            cotracker = CoTrackerThreeOnline(
+                stride=4, corr_radius=3, window_len=window_len
+            )
+
+    if checkpoint is not None:
+        with open(checkpoint, "rb") as f:
+            state_dict = torch.load(f, map_location="cpu")
+            if "model" in state_dict:
+                state_dict = state_dict["model"]
+        cotracker.load_state_dict(state_dict)
+    return cotracker

torch_hub/facebookresearch_co-tracker_main/cotracker/models/core/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.

torch_hub/facebookresearch_co-tracker_main/cotracker/models/core/cotracker/__init__.py

@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.