Awesome

Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation

📄[arXiv]   📄[PDF]

This repository contains code for CVPR2024 paper:

Decoupling Static and Hierarchical Motion Perception for Referring Video Segmentation
Shuting He, Henghui Ding
CVPR 2024

Installation:

Please see INSTALL.md. Then

pip install -r requirements.txt
python3 -m spacy download en_core_web_sm

Inference

1. Val^u set

Obtain the output masks of Val^u set:

python train_net_dshmp.py \
    --config-file configs/dshmp_swin_tiny.yaml \
    --num-gpus 8 --dist-url auto --eval-only \
    MODEL.WEIGHTS [path_to_weights] \
    OUTPUT_DIR [output_dir]

Obtain the J&F results on Val^u set:

python tools/eval_mevis.py

2. Val set

Obtain the output masks of Val set for CodaLab online evaluation:

python train_net_dshmp.py \
    --config-file configs/dshmp_swin_tiny.yaml \
    --num-gpus 8 --dist-url auto --eval-only \
    MODEL.WEIGHTS [path_to_weights] \
    OUTPUT_DIR [output_dir] DATASETS.TEST '("mevis_test",)'

Training

Firstly, download the backbone weights (model_final_86143f.pkl) and convert it using the script:

wget https://dl.fbaipublicfiles.com/maskformer/mask2former/coco/instance/maskformer2_swin_tiny_bs16_50ep/model_final_86143f.pkl
python tools/process_ckpt.py
python tools/get_refer_id.py

Then start training:

python train_net_dshmp.py \
    --config-file configs/dshmp_swin_tiny.yaml \
    --num-gpus 8 --dist-url auto \
    MODEL.WEIGHTS [path_to_weights] \
    OUTPUT_DIR [path_to_weights]

Note: We train on a 3090 machine using 8 cards with 1 sample on each card, taking about 17 hours.

Models

☁️ Google Drive

Acknowledgement

This project is based on MeViS. Many thanks to the authors for their great works!

BibTeX

Please consider to cite DsHmp if it helps your research.

@inproceedings{DsHmp,
  title={Decoupling static and hierarchical motion perception for referring video segmentation},
  author={He, Shuting and Ding, Henghui},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={13332--13341},
  year={2024}
}