We provide the code for reproducing experiment results of SCNet.
- SCNet addresses the limitation in training-inference sample distribution mismatch of existing cascade method for instance segmentation.
- SCNet strengen the relationship between subtasks: classification, detection and segmentation.
- State-of-the-art: without bell and whistle, SCNet achieves 44.7 box AP and 42.3 mask AP on ResNext-101. This can be further improved with well-known plugins and tricks, such as Group Norm, DCN, multi-scale training/testing.
- Fast training and testing: SCNet achieves better performancce while training/inference faster and requires less memory compared to Cascade Mask R-CNN and HTC. See comparison below.
SCNet requires COCO and COCO-stuff dataset for training. You need to download and extract it in the COCO dataset path. The directory should be like this.
mmdetection
├── mmdet
├── tools
├── configs
├── data
│ ├── coco
│ │ ├── annotations
│ │ ├── train2017
│ │ ├── val2017
│ │ ├── test2017
| | ├── stuffthingmaps
##Results and Models
The results on COCO 2017val are shown in the below table. (results on test-dev are usually slightly higher than val)
| Backbone | Style | Lr schd | Mem (GB) | Inf speed (fps) | box AP | mask AP | TTA box AP | TTA mask AP | Config | Download |
|---|---|---|---|---|---|---|---|---|---|---|
| R-50-FPN | pytorch | 1x | 7.0 | 6.2 | 43.5 | 39.2 | 44.8 | 40.9 | config | model | log |
| R-50-FPN | pytorch | 20e | 7.0 | 6.2 | 44.5 | 40.0 | 45.8 | 41.5 | config | model | log |
| R-101-FPN | pytorch | 20e | 8.9 | 5.8 | 45.8 | 40.9 | 47.3 | 42.7 | config | model | log |
| X-101-64x4d-FPN | pytorch | 20e | 13.2 | 4.9 | 47.5 | 42.3 | 48.9 | 44.0 | config | model | log |
- Training hyper-parameters are identical to those of HTC.
- TTA means Test Time Augmentation, which applies horizonal flip and multi-scale testing. Refer to config.
If you find our work helpful for your research. Please cite our paper.
@inproceedings{vu2019cascade,
title={SCNet: Training Inference Sample Consistency for Instance Segmentation},
author={Vu, Thang and Haeyong, Kang and Yoo, Chang D},
booktitle={AAAI},
year={2021}
}