DCT-Mask: Discrete Cosine Transform Mask Representation for Instance Segmentation
This project hosts the code for implementing the DCT-MASK algorithms for instance segmentation.
[DCT-Mask: Discrete Cosine Transform Mask Representation for Instance Segmentation] Xing Shen*, Jirui Yang*, Chunbo Wei, Bing Deng, Jianqiang Huang, Xiansheng Hua Xiaoliang Cheng, Kewei Liang
In Proceedings of the IEEE conference on Computer Vision and Pattern Recognition(CVPR 2021)
arXiv preprint(arXiv:2011.09876)
Contributions
- We propose a high-quality and low-complexity mask representation for instance segmentation, which encodes the high-resolution binary mask into a compact vector with discrete cosine transform.
- With slight modifications, DCT-Mask could be integrated into most pixel-based frameworks, and achieve significant and consistent improvement on different datasets, backbones, and training schedules. Specifically, it obtains more improvements for more complex backbones and higher-quality annotations.
- DCT-Mask does not require extra pre-processing or pre-training. It achieves high-resolution mask prediction at a speed similar to low-resolution.
Installation
Requirements
- PyTorch ≥ 1.5 and fvcore == 0.1.1.post20200716
This implementation is based on detectron2. Please refer to INSTALL.md. for installation and dataset preparation.
Usage
The codes of this project is on projects/DCT_Mask/
Train with multiple GPUs
cd ./projects/DCT_Mask/
./train1.sh
Testing
cd ./projects/DCT_Mask/
./test1.sh
Model ZOO
Trained models on COCO
| Model | Backbone | Schedule | Multi-scale training | Inference time (s/im) | AP (minival) | Link |
|---|---|---|---|---|---|---|
| DCT-Mask R-CNN | R50 | 1x | Yes | 0.0465 | 36.5 | download(Fetch code: xpdm) |
| DCT-Mask R-CNN | R101 | 3x | Yes | 0.0595 | 39.9 | download(Fetch code: 7q6x) |
| DCT-Mask R-CNN | RX101 | 3x | Yes | 0.1049 | 41.2 | download(Fetch code: ufw2) |
| Casecade DCT-Mask R-CNN | R50 | 1x | Yes | 0.0630 | 37.5 | download(Fetch code: yqxp) |
| Casecade DCT-Mask R-CNN | R101 | 3x | Yes | 0.0750 | 40.8 | download(Fetch code: r8xv) |
| Casecade DCT-Mask R-CNN | RX101 | 3x | Yes | 0.1195 | 42.0 | download(Fetch code: pdej) |
Trained models on Cityscapes
| Model | Data | Backbone | Schedule | Multi-scale training | AP (val) | Link |
|---|---|---|---|---|---|---|
| DCT-Mask R-CNN | Fine-Only | R50 | 1x | Yes | 37.0 | download(Fetch code: dn7i) |
| DCT-Mask R-CNN | CoCo-Pretrain +Fine | R50 | 1x | Yes | 39.6 | download(Fetch code: ntqf) |
Notes
- We observe about 0.2 AP noise in COCO.
- High variance observed in CityScapes when trained on fine annotations only. We report the median of 5 runs AP in the article (i.e. 35.6), while in this repo we report the best results (37.0).
- Initialized from COCO pre-training will reduce the variance on CityScapes as well as increasing mask AP.
- The inference time is measured on single GPU with batchsize 1. All GPUs are NVIDIA V100.
- Lvis 0.5 is used for evaluation.
Contributing to the project
Any pull requests or issues are welcome.
If there is any problem with this project, please contact Xing Shen.
Citations
Please consider citing our papers in your publications if the project helps your research.
License
- MIT License.
168 Dec 29, 2022
11 Dec 30, 2022
169 Nov 29, 2022
21k Jan 06, 2023
12 Nov 11, 2022
14 Apr 05, 2022
1k Jan 02, 2023
4 Feb 13, 2022
13 Jul 25, 2022
593 Jan 03, 2023
2 Jul 12, 2022
605 Jan 03, 2023
9 Nov 29, 2022
50 Nov 11, 2022
15 Oct 24, 2022
575 Dec 06, 2022
648 Jan 08, 2023
77 Nov 22, 2022
6 Dec 12, 2022
27 Nov 05, 2022