MoSA Checkpoints

This repository provides the pretrained checkpoints for MoSA (Motion-Grounded Segment Anything), introduced in Seeing as Humans Do: Learning from Motion to Segment Anything Without Supervision (ECCV 2026).

Code and usage instructions are available in the MoSA GitHub repository. The paper link will be added after the official release.

Checkpoints

File	Backbone	Purpose
pgm.pth	ViT-B/8	Pseudo-mask generation
promptable_segmentation.pth	Swin-Tiny	Point-promptable segmentation
whole_image_segmentation.pth	ResNet-50	Whole-image segmentation
moby_swin_t_300ep_pretrained.pth	Swin-Tiny	Third-party model initialization

Third-party checkpoint: moby_swin_t_300ep_pretrained.pth is the MoBY Swin-Tiny model pretrained for 300 epochs on ImageNet-1K. It originates from the official Transformer-SSL MoBY release and is used by the UnSAM configuration. It was not trained by the MoSA authors; please follow the upstream terms and citation requirements.

Usage

Please follow the training and evaluation instructions in the MoSA repository. The moby_swin_t_300ep_pretrained.pth file is only an initialization checkpoint; use the other three files for MoSA inference and evaluation.

Citation

The paper and BibTeX entry will be added after the official ECCV 2026 release.

Acknowledgements

The adaptation pipeline is based on UnSAM. The Swin-Tiny initialization checkpoint is provided by the official MoBY project.