| --- |
| license: apache-2.0 |
| tags: |
| - point-tracking |
| - onnx |
| - causal |
| --- |
| # Causal BootsTAPIR: 256px, three query points |
| Official Google causal BootsTAPIR weights converted in GitHub Actions into encoder, query and recurrent-step ONNX graphs. |
| This is point tracking from three supplied initial image coordinates, not semantic foot detection, 3D reconstruction, or independent toe articulation. |
| Source code: https://github.com/google-deepmind/tapnet/tree/730cda1c730877cfedbe01bf87fb1cadb78a565d |
| Source weights: https://huggingface.co/google/tapnet/resolve/5d3fb48e76c5422841e38501514121e251beabb7/causal_bootstapir_checkpoint.pt |
| Source SHA-256: `87c1e752cf5ce56e3e2f7da460aeb4d40fc826d04ef2939bade86a5c7495377f`. |
| Input: RGB float32 `[1,1,256,256,3]` in `[-1,1]`; first-frame queries `[1,3,2]` in model-pixel `(y,x)` order. Outputs are xy pixel tracks and two uncertainty logits. |
| Run encoder then query once; run encoder and step for each frame, retaining query features and feeding `new_causal_state` into the next `causal_state`. Start from the zero state defined by `tapir-models.json`. |
| The first-frame query extraction uses a mathematically equivalent 2D sample because time is fixed at zero. It is checked against the official 5D implementation. |
| The Actions comparison uses only a public CC BY image with artificial motion/occlusion. It verifies conversion numerics, not real-foot tracking quality or occlusion recovery accuracy. |
| WebGPU operator coverage is checked statically against ONNX Runtime Web 1.22.0; GPU execution, latency and RTX memory use are not measured by the CPU comparison. |
| No private video, extracted private frame, avatar file, or private pose trace is included. |
| License: Apache-2.0. Copyright and source notices are retained in LICENSE and the pinned source links. |
|
|
