Source code for ICCV-25 paper 'Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation'
EndoVis-18 - All Test Videos - Inst. w/ Tis. Segmentation (J&F / J / F) - Zero-Shot
| Methods | J&F | J | F |
|---|---|---|---|
| Ours | 74.6 | 76.1 | 73.1 |
| Baseline | 73.3 | 75.1 | 71.6 |
EndoVis-18 - SEQ 15 - Tissue Segmentation (J&F / J / F) - Zero-Shot
| Obj | Ours | Baseline |
|---|---|---|
| 011 | 36.5 / 32.7 / 40.2 | 33.2 / 29.8 / 36.6 |
| 012 | 68.0 / 89.3 / 46.7 | 64.2 / 86.9 / 41.5 |
| 017 | 82.3 / 90.4 / 74.1 | 81.2 / 89.6 / 72.8 |
Figures are arranged in a 2×2 grid: top-left Image, bottom-left GT, top-right Baseline, and bottom-right Ours.
oasis-endovis-s15-fullvid.webm
- [10/2025] Repo Release
- [08/2025] Sorry for busy chasing other conferences. The code is now being cleaned and will be make public.
- [07/2025] We released our work 'OASIS', the paper is now on Arxiv.
- More checkpoints and results on surgical videos coming in...
- Checkpoints & Pre-computed results...
- Training & Inference Code release
- Initialization
- Python
- PyTorch
By check the ckpts/README.md and finish the download of datasets and image-pretrained ckpts, could leverage the train.sh to start model training. Note that u may want to activate the environment before run the script.
CUDA_VISIBLE_DEVICES=0,1,2,3 OMP_NUM_THREADS=4 torchrun \
--master_port 12345 \
--nproc_per_node=4 \
oasis/train.py \
exp_id=main_small \
model=small \ # Model size/version
data=davis # Training datasetsIf you find this project helpful in your research, please consider citing our papers:
@inproceedings{qin2025structure,
title={Structure Matters: Revisiting Boundary Refinement in Video Object Segmentation},
author={Qin, Guanyi and Wang, Ziyue and Shen, Daiyun and Liu, Haofeng and Zhou, Hantao and Wu, Junde and Hu, Runze and Jin, Yueming},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
month={October},
year={2025}
}
We borrowed some parts from the following open-source projects:
Special thanks to them.