This is a detectron2 version of SeqFormer, which supports multi-card training and inference, and supports inference by clip-matching manner. Joint-training with the COCO is not supported yet, and this will be added in a future version. You can refer to the original code of SeqFormer for more details about joint-training.
Train on YouTube-VIS 2019, evaluate on YouTube-VIS 2019.
| Name | AP | AP50 | AP75 | AR1 | AR10 | download |
|---|---|---|---|---|---|---|
| R50 | ||||||
| R101 | ||||||
| SwinL |
Model Zoo of SeqFormer is in preparation.
We performed the experiment on NVIDIA Tesla V100 GPU. All models of SeqFormer are trained with total batch size of 32.
To train SeqFormer on YouTube-VIS 2019 with 8 GPUs , run:
python3 projects/SeqFormer/train_net.py --config-file projects/SeqFormer/configs/XXX.yaml --num-gpus 8
Evaluating on YouTube-VIS 2019:
python3 projects/SeqFormer/train_net.py --config-file projects/SeqFormer/configs/XXX.yaml --num-gpus 8 --eval-only
Evaluating on YouTube-VIS 2019 by a clip-matching manner:
python3 projects/SeqFormer/train_net.py --config-file projects/SeqFormer/configs/XXX.yaml --num-gpus 8 --eval-only MODEL.SeqFormer.CLIP_MATCHING True MODEL.SeqFormer.CLIP_LENGTH $CLIP_LENGTH MODEL.SeqFormer.CLIP_STRIDE $CLIP_STRIDE
To get quantitative results, please zip the json file and upload to the codalab server.
@inproceedings{seqformer,
title={SeqFormer: Sequential Transformer for Video Instance Segmentation},
author={Wu, Junfeng and Jiang, Yi and Bai, Song and Zhang, Wenqing and Bai, Xiang},
booktitle={ECCV},
year={2022},
}
This repo is based on detectron2, Deformable DETR, VisTR, and IFC Thanks for their wonderful works.