[PatchDCT: Patch Refinement for High Quality Instance Segmentation] Qinrou Wen, Jirui Yang, Xue Yang, Kewei Liang
arXiv preprint(arXiv:2302.02693)
In this repository, we release code for PatchDCT in Detectron2.
- PatchDCT is the fist compressed vector based multi-stage refinement framework.
- By using a classifier to refine foreground and background patches, and predicting an informative low-dimensional DCT vector for each mixed patch, PatchDCT generates high-resolution masks with fine boundaries and low computational cost.
- PyTorch ≥ 1.8
This implementation is based on detectron2. Please refer to INSTALL.md. for installation and dataset preparation.
The codes of this project is on projects/PatchDCT/
cd ./projects/PatchDCT/
./train.sh
cd ./projects/PatchDCT/
./test.sh
cd ./projects/PatchDCT/
./test_speed.sh
cd ./projects/PatchDCT/
./test_up.sh
For Swin-B backbone, use train_net_swinb.py instead of train_net.py
Model | Backbone | Schedule | Multi-scale training | FPS | AP (val) | Link |
---|---|---|---|---|---|---|
PatchDCT | R50 | 1x | Yes | 12.3 | 37.2 | download |
PatchDCT | R101 | 3x | Yes | 11.8 | 40.5 | download |
PatchDCT | RX101 | 3x | Yes | 11.7 | 41.8 | download |
PatchDCT | SwinB | 3x | Yes | 7.3 | 46.1 | download |
Model | Data | Backbone | Schedule | Multi-scale training | AP (val) | Link |
---|---|---|---|---|---|---|
PatchDCT | Fine-Only | R50 | 1x | Yes | 38.2 | download |
PatchDCT | COCO Pretrain+Fine | R50 | 1x | Yes | 40.3 | download |
- We observe about 0.2 AP noise in COCO.
- The inference time is measured on NVIDIA A100 with batchsize=1.
- Lvis 0.5 is used for evaluation.
Any pull requests or issues are welcome.
If there is any problem with this project, please contact Qinrou Wen.
Please consider citing our papers in your publications if the project helps your research.
- MIT License.