Add four-view motion render previews on Y Bot (MP4 + camera JSON per view) and document them

.gitattributes

@@ -1 +1,3 @@
 *.fbx filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -2,3 +2,7 @@
 __pycache__/
 *.pyc
 .DS_Store
+
+# Per-frame render PNGs — local only; the Hub ships per-view MP4s
+animation_motion_render/*/v[0-9][0-9][0-9]/
+animation_motion_render/*/frames.tar

README.md

@@ -34,10 +34,9 @@ A complete snapshot of the [Mixamo](https://www.mixamo.com/) library: **2,317 mo
 All animations share one uniform 65-joint `mixamorig` skeleton, so any clip can drive any
 compatible character without remapping.
 
-**Use `animation_motion/`.** The full export contains 2,446 files, but 129 of them are single-pose
-assets that carry no motion at all — Mixamo's `*_Pose*` library, plus `T-Pose` and `Leaning`.
-`animation_motion/` is the 2,317 that actually move; `animation/` keeps all 2,446 for anyone who
-wants the poses too. See [Static clips](#static-clips) for how they were separated.
+**Use `animation_motion/` and `character_refined/`.** The full export contains 2,446 animation
+files, but 129 are single-pose assets that carry no motion (Mixamo's `*_Pose*` library, plus
+`T-Pose` and `Leaning`); `animation_motion/` is the 2,317 that actually move.
 
 ## Dataset structure
 
@@ -46,6 +45,10 @@ animation_motion/           2,317 .fbx   clips that move  <- use this
 animation/                  2,446 .fbx   the full export, poses included
 character/                    114 .fbx   T-pose characters with skin mesh, as exported
 character_refined/            114 .fbx   same characters, bone namespaces normalised  <- use this
+character_refined_glb/        114 .glb   character_refined/ converted to GLB, same stems
+character_tpose/              114 .png   rest-pose render grid per character
+animation_motion_render/    2,317 dirs   four-view render of every moving clip on Y Bot
+                                         (MP4 + camera JSON per view)
 
 animation_motion_prompts.json            filename -> text label + Mixamo id, 2,317 entries
 animation_motion_frames.json             filename -> keyframe count, 2,317 entries
@@ -69,36 +72,15 @@ Files in `animation_motion/` are the same bytes as their counterparts in `animat
 
 ## Static clips
 
-129 of the 2,446 exported files hold a pose rather than a motion. They are not a download error —
-Mixamo ships a `*_Pose*` library (`Male_Sitting_Pose_12`, `Female_Standing_Pose_14`, …) alongside
-its animations, and those exports are two frames of the same pose.
-
-They were found by decoding every animation curve and measuring its peak-to-peak range, not by
-filename — keyframe count cannot tell them apart, since a static pose still writes a full set of
-keys, and two of them (`T-Pose`, `Leaning`) are not named like poses at all.
-
-`static_clips.csv` lists all 129 with the measurement and the reason:
-
-| reason | count | |
-|---|---|---|
-| `all curves flat` | 122 | every curve identical frame to frame |
-| `float32 rounding noise only` | 6 | peak-to-peak 0.001–0.024, four orders of magnitude below any real clip |
-| `two-frame pose delta, no in-between frames` | 1 | `Female_Sitting_Pose_7`: a real pose change, but only two frames |
-
-The split is unambiguous. The largest static reading is 0.024 and the smallest moving one is 1.115
-(`Sitting_Idle_4`, a 90-frame breathing idle); nothing falls between. Frame counts separate them
-too — every static file has 2 frames, while the shortest real clip has 8.
-
-Quiet clips are **not** filtered: `Idle_5`, `Rifle_Kneel_Idle`, `Sword_And_Shield_Block_Idle` and
-similar read 1–3 peak-to-peak and are kept, because a breathing idle or a held guard is a motion.
+The 129 pose files were found by decoding every animation curve and measuring its peak-to-peak
+range, not by filename — the split is unambiguous (largest static reading 0.024, smallest moving
+one 1.115, nothing between). Quiet clips like breathing idles and held guards are kept.
+`static_clips.csv` lists all 129 with the measurement and the reason.
 
 ## Tabular index
 
-Two CSVs make the dataset browsable in the dataset viewer and loadable without downloading
-any FBX. They are generated from the JSON indexes by `scripts/build_tables.py`.
-
-`metadata.csv` — one row per animation. It covers all 2,446 files; join against
-`static_clips.csv` to restrict it to the ones that move:
+Two CSVs make the dataset browsable in the dataset viewer and loadable without downloading any
+FBX. `metadata.csv` — one row per animation, covering all 2,446 files:
 
 | column | meaning |
 |---|---|
@@ -109,13 +91,6 @@ any FBX. They are generated from the JSON indexes by `scripts/build_tables.py`.
 | `frames` | keyframe count |
 | `motion_id` | Mixamo motion UUID |
 
-```python
-import csv
-static = {r["file"] for r in csv.DictReader(open("static_clips.csv"))}
-moving = [r for r in csv.DictReader(open("metadata.csv"))
-          if r["file"].split("/")[-1] not in static]      # 2,317 rows
-```
-
 `characters.csv` — one row per character, with `bone_count` and `covers_animation_rig`
 (`true` for the 97 characters whose skeleton covers all 65 animation bones).
 
@@ -125,18 +100,8 @@ from datasets import load_dataset
 anims = load_dataset("Linzhan/Mixamo-Animations-Characters", "animations", split="train")
 chars = load_dataset("Linzhan/Mixamo-Animations-Characters", "characters", split="train")
 
-# clips between 1 and 10 seconds at 30 fps
-mid = anims.filter(lambda r: 30 <= r["frames"] <= 300)
-
-# characters that any animation can drive
-ok = chars.filter(lambda r: r["covers_animation_rig"] == "true")
-```
-
-Fetch only the FBX you need:
-
-```python
-from huggingface_hub import hf_hub_download
-p = hf_hub_download("Linzhan/Mixamo-Animations-Characters", "animation_motion/Jab_Cross.fbx", repo_type="dataset")
+mid = anims.filter(lambda r: 30 <= r["frames"] <= 300)                 # 1–10 s at 30 fps
+ok  = chars.filter(lambda r: r["covers_animation_rig"] == "true")      # drop-in retargetable
 ```
 
 ## Animations
@@ -144,51 +109,68 @@ p = hf_hub_download("Linzhan/Mixamo-Animations-Characters", "animation_motion/Ja
 - **Count**: 2,317 clips that move (2,446 including poses)
 - **Format**: binary FBX 7.7, 30 fps, skeleton + animation curves only (no mesh)
 - **Skeleton**: identical 65-joint `mixamorig` rig in *every* clip
-- **Length**: 8 – 1,884 keyframes in `animation_motion/` (2 – 1,884 across the full export)
+- **Length**: 8 – 1,884 keyframes in `animation_motion/`
 
-Files are named after the motion, e.g. `Jab_Cross.fbx`. Distinct motions that share a Mixamo
-name are disambiguated with a numeric suffix (`Waving.fbx`, `Waving_2.fbx`, `Waving_3.fbx`).
+Files are named after the motion, e.g. `Jab_Cross.fbx`; distinct motions sharing a Mixamo name
+get a numeric suffix (`Waving.fbx`, `Waving_2.fbx`).
 
 ### `animation_motion_prompts.json`
 
 ```json
 "180_Turn_W_Briefcase.fbx": {
   "prompt": "180 Turn W/ Briefcase",
-  "description": "Female 180 Turn With Briefcase",
+  "description": "180 Turn With Briefcase",
   "motion_id": "c9c9f110-b96c-11e4-a802-0aaa78deedf9"
 }
 ```
 
 | field | meaning |
 |---|---|
-| `prompt` | Mixamo's catalogue name. Present for every clip. Short (mean 2.6 words), sometimes abbreviated (`W/`, `2hand`, `L`/`R`). |
-| `description` | Mixamo's longer description. Present for 2,395 clips (97.9%). Fuller natural language (mean 5.0 words) and usually more specific. |
-| `motion_id` | Mixamo's internal motion UUID. |
-
-`description` is only stored when it differs from `prompt`.
+| `prompt` | Mixamo's catalogue name. Present for every clip; short (mean 2.6 words), sometimes abbreviated (`W/`, `2hand`, `L`/`R`) |
+| `description` | Mixamo's longer description; fuller and usually more specific. Present for every entry — filled with `prompt` where Mixamo ships none, so it can be read without a fallback |
+| `motion_id` | Mixamo's internal motion UUID |
 
-For text-conditioned tasks, `description or prompt` is the more discriminative label:
-`prompt` alone yields 1,562 unique strings across 2,446 clips, while preferring `description`
-yields 2,279. Example — `"Angry Gesture"` covers three different motions, separated only by
-their descriptions (`Angrily Pumping Fists Forward`, `Seated Idle With Hands On A Table Being
-Rude`, `Angry Forward Gesture`).
+For text-conditioned tasks, `description or prompt` is the more discriminative label
+(2,134 unique strings vs 1,541 from `prompt` alone).
 
 ### `animation_motion_frames.json`
 
-```json
-"180_Turn_W_Briefcase.fbx": 68
-```
-
-Keyframe count parsed from each FBX (`scripts/fbx_frames.py`), not a nominal duration.
+`"180_Turn_W_Briefcase.fbx": 68` — keyframe count parsed from each FBX, not a nominal duration.
 
 ### `animation_bones.json`
 
-```json
-{ "mixamo": ["mixamorig:Head", "mixamorig:HeadTop_End", "mixamorig:Hips", ...] }
+`{ "mixamo": ["mixamorig:Head", "mixamorig:HeadTop_End", "mixamorig:Hips", ...] }` — the 65
+bone names present in every animation: hips/spine/neck/head chain, both arms with full
+four-joint fingers, and both legs with toes.
+
+## Renders (`animation_motion_render/`)
+
+A four-view render of every clip in `animation_motion/`, driven on
+`character_refined/Y_Bot.fbx` — the character whose skeleton the animations are authored on,
+so no retargeting touches the motion. One folder per clip, named after the FBX stem:
+
+```
+animation_motion_render/Jab_Cross/
+  v000.mp4  v000.json      front  (azimuth  -90°)
+  v001.mp4  v001.json      back   (azimuth  +90°)
+  v002.mp4  v002.json      left   (azimuth  180°)
+  v003.mp4  v003.json      right  (azimuth    0°)
 ```
 
-The 65 bone names present in every animation: hips/spine/neck/head chain, both arms with full
-four-joint fingers, and both legs with toes.
+Rendered with Blender EEVEE at 512×512, 30 fps, elevation 0, and composed into one H.264 MP4
+per view — the transparent-film frames flattened onto a light-grey background, since H.264
+carries no alpha. The MP4s preview each motion directly on the Hub; the per-frame RGBA PNGs
+behind them are not hosted (they would add over 800k files), so extract stills from the MP4s
+(`ffmpeg -i v000.mp4 frames/%04d.png`) if you need them.
+
+Rendering covers the first 200 frames of a clip (~6.7 s); 375 of the 2,317 clips are longer
+and are truncated — `frame_start`/`frame_end` in the per-view JSON record the covered range.
+Frame counts otherwise match `animation_motion_frames.json` exactly.
+
+Each `v00X.json` holds the camera in NeRF convention: `view`, `azimuth`/`elevation` (radians),
+`resolution`, `camera_angle_x` (horizontal FoV, radians), the 4×4 camera-to-world
+`transform_matrix`, the `scene_scale`/`scene_offset` that normalised the character into frame,
+and `frame_start`/`frame_end`/`fps`.
 
 ## Characters
 
@@ -201,40 +183,31 @@ actual bone names per character.
 
 ### `character/` vs `character_refined/`
 
-27 of the 114 characters export with a **numbered bone namespace** — `mixamorig1:Hips`,
-`mixamorig7:Hips`, `mixamorig12:Hips` — instead of `mixamorig:Hips`. Those names do not match
-the animations, so retargeting silently fails.
+27 of the 114 characters export with a **numbered bone namespace** — `mixamorig1:Hips` instead
+of `mixamorig:Hips` — which does not match the animations, so retargeting silently fails.
+`character/` keeps the files exactly as exported; `character_refined/` renames the namespace
+back to `mixamorig:` with a surgical binary-FBX rewrite that leaves every non-string property
+byte-identical (verified both ways by `scripts/verify_refined.py`). **Use `character_refined/`.**
+The affected files are listed in `character_bone_fix_applied.txt`.
 
-- `character/` keeps the files exactly as exported, unmodified.
-- `character_refined/` renames the namespace back to `mixamorig:`. **Use this directory.**
+### `character_refined_glb/`
 
-The rename is a surgical binary-FBX rewrite (`scripts/fbx_rename_bones.py`): only string
-properties change, and mesh, geometry, materials and textures are byte-identical to the
-originals. This is verified two ways — a no-edit round-trip of the writer reproduces its input
-byte-for-byte, and `scripts/verify_refined.py` walks both node trees in parallel to confirm
-every non-string property is unchanged and every bone name is preserved 1:1.
+The same 114 characters converted one-to-one to GLB — mesh, rig and materials with textures
+embedded, no animation — for toolchains that read glTF rather than FBX. Same stems as the FBXs.
 
-The affected files are listed in `character_bone_fix_applied.txt`.
+### `character_tpose/`
 
-## Retargeting compatibility
+One 1024×1024 PNG per character, named after the FBX stem: a 2×2 grid of its rest pose —
+front and back on the top row, left and right below — rendered (EEVEE) from
+`character_refined/`. A visual index of the 114 characters.
 
-Bone names are read from the FBX node tree (`scripts/fbx_bones.py`).
+## Retargeting compatibility
 
 | | count | notes |
 |---|---|---|
 | Cover all 65 animation bones | **97 / 114** | drop-in retargetable |
-| Reduced `mixamorig` rig | 13 | 37–64 bones, nearly always fewer finger joints (e.g. Mutant 37, Castle_Guard_01 43, Aciana 57). Retargeting works; absent bones simply receive no data. |
-| Different bone naming | 4 | listed in `character_nonstandard_rig.txt` |
-
-The four exceptions are mostly trivial to map, since three of them use standard Mixamo bone
-names with a different prefix:
-
-| character | bones | naming |
-|---|---|---|
-| `Alexis` | 77 | `Alexis:Head` — standard names, custom prefix |
-| `Alpha` | 53 | `Head` — standard names, no prefix |
-| `Wizard` | 67 | `Head` — standard names, no prefix |
-| `Aiden` | 50 | `aiden:Bip01 Head` — 3ds Max Biped naming, needs a real mapping |
+| Reduced `mixamorig` rig | 13 | 37–64 bones, nearly always fewer finger joints. Retargeting works; absent bones simply receive no data |
+| Different bone naming | 4 | listed in `character_nonstandard_rig.txt`; three use standard names with a different prefix, one (`Aiden`) uses 3ds Max Biped naming |
 
 ## Usage
 
@@ -262,27 +235,22 @@ bpy.ops.import_scene.fbx(filepath="animation_motion/Jab_Cross.fbx")
 
 ## Known limitations
 
-- **129 of the 2,446 exported files hold a pose, not a motion**, and are excluded from
-  `animation_motion/`. See [Static clips](#static-clips).
-- **51 clips have no `description`** (2.1%). 36 of them have a unique `prompt`, so the
-  `description or prompt` fallback is unambiguous. The other 15 share a `prompt` with other
-  clips and therefore cannot be told apart from text alone — the worst case is
-  `Hit_Reaction_4.fbx`, one of nine clips named `Hit Reaction`.
-- Even preferring `description`, 167 clips still share a label with another clip.
-- `description` frequently carries a `Female`/`Male` prefix. This is animation style rather
-  than motion semantics; strip it if that is not a signal you want.
-- 126 descriptions are comma-separated lists or carry a trailing ` - ` production note.
+- **62 clips have no `description`**; 23 of those share a `prompt` with other clips and cannot
+  be told apart from text alone. Even preferring `description`, 341 clips share a label with
+  another clip.
+- In `animation_motion_prompts.json`, `Female`/`Male` style prefixes and leading
+  `Sequence -` markers have been stripped from `prompt`/`description` — they describe
+  animation style, not motion semantics. `animation_prompts.json`, `metadata.csv` and the
+  filenames keep the original Mixamo labels.
 - Mixamo `MotionPack` bundles are excluded, as their constituent clips are present individually.
 
 ## Provenance
 
-Exported from Mixamo through its own export API with `format=fbx7_2019`, `skin=false` for
-animations and `skin=true` for characters, retargeted onto the standard Y Bot skeleton.
-
-Integrity was verified end to end: every FBX parses without truncation, no byte-identical
-duplicates exist, all index keys match the files exactly, every frame count was re-derived from
-the FBX, every animation curve was decoded to separate motion from pose, and all 2,446 motion
-names/ids plus 114 character names were cross-checked against the Mixamo API. The `scripts/` directory reproduces the download, the indexes and all checks.
+Exported from Mixamo through its own export API (`format=fbx7_2019`, `skin=false` for
+animations, `skin=true` for characters), retargeted onto the standard Y Bot skeleton. Every FBX
+parses without truncation, every index key matches its file, and all names/ids were
+cross-checked against the Mixamo API; `scripts/` reproduces the download, the indexes and all
+checks.
 
 ## Terms