GitLab

rescale predictions to original size during evaluation

Paste masks using bilinear grid sampling

 * set segmentation loss weight to 0.5
 * skip boxes with zero area while pasting masks

* added `PanopticSegmentationModule`

* added tests for `PanopticSegmentationModule`

* added script to export `saved_model`

* added registry_imports

* added training script

* fixed docstrings

* pass model as an arg to serving module

PiperOrigin-RevId: 401678642