PRIMA: Boosting Animal Mesh Recovery with Biological Priors and Test-Time Adaptation

Upload an animal image. The demo runs Detectron2 for animal detection, PRIMA for 3D pose/shape, DeepLabCut SuperAnimal for 2D keypoints, and test-time adaptation (TTA) with configurable learning rate and iterations. Set TTA iterations to 0 to disable adaptation.

Results (PNG/OBJ and 26-keypoint visualizations) are saved under 'demo_out_tta_gradio'.

Input image

TTA learning rate

1e-7 0.0001

TTA iterations

Set to 0 to disable TTA and reuse the initial PRIMA prediction.

0 100

Detection threshold

0.3 0.9

Keypoint confidence threshold

0 1

Render side view

Save meshes (.obj)

Before TTA

After TTA

PRIMA 26 keypoints

Examples

Input image	TTA learning rate	TTA iterations	Detection threshold	Keypoint confidence threshold	Render side view	Save meshes (.obj)