Pose Tracking
Collection
This collections hosts a set of pose tracking models developed at the Allen Institute • 1 item • Updated • 1
Lightning Pose Model
Trained by Anna Lakunina on DATA_NAME (the S3 path or data asset ID).
- Backbone:
resnet50_animal_ap10k - Losses used:
['pca_singleview'] - Keypoints (18):
eye_top, eye_right, eye_bottom, eye_left, pupil_top, pupil_right, pupil_bottom, pupil_left, pupil_top_right, pupil_bottom_right, pupil_bottom_left, pupil_top_left, reflection, nostril, nose_tip, whisker_1, whisker_2, whisker_3 - Epochs: 300-300, train/val split: 0.95/0.05
- Supervised RMSE (validation): 0.64 resized-px
Performance
Supervised RMSE (from TensorBoard logs, unit: resized-px - measured on the resized training image, not original-frame pixels). Lower is better.
No per-keypoint pixel-error file was found; the numbers below come from the training logs instead.
Source: events.out.tfevents.1736539402.f2768a15001d.48.0
| Set | Final RMSE | Best RMSE |
|---|---|---|
| train | 0.34 | 0.29 |
| validation | 0.77 | 0.64 |
Changelog
v1.0
what this model is / what change
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support