This repo is the official implementation of RefineMask: Towards High-Quality Instance Segmentation with Fine-Grained Features.
Update: A faster and slightly better implementation is available! You can train R50-RefineMask 2 images per gpu with less than 11G memory cost now. See here.
Method | Backbone | Schedule | AP | AP* | Checkpoint |
---|---|---|---|---|---|
Mask R-CNN | R50-FPN | 1x | 34.7 | 36.8 | |
RefineMask | R50-FPN | 1x | 37.3 | 40.6 | download |
Mask R-CNN | R50-FPN | 2x | 35.4 | 37.7 | |
RefineMask | R50-FPN | 2x | 37.8 | 41.2 | download |
Mask R-CNN | R101-FPN | 1x | 36.1 | 38.4 | |
RefineMask | R101-FPN | 1x | 38.6 | 41.8 | download |
Mask R-CNN | R101-FPN | 2x | 36.6 | 39.3 | |
RefineMask | R101-FPN | 2x | 39.0 | 42.4 | download |
Note: No data augmentations except standard horizontal flipping were used.
Method | Backbone | Schedule | AP | APr | APc | APf | Checkpoint |
---|---|---|---|---|---|---|---|
Mask R-CNN | R50-FPN | 1x | 22.1 | 10.1 | 21.7 | 30.0 | |
RefineMask | R50-FPN | 1x | 25.7 | 13.8 | 24.9 | 31.8 | download |
Mask R-CNN | R101-FPN | 1x | 23.7 | 12.3 | 23.2 | 29.1 | |
RefineMask | R101-FPN | 1x | 27.1 | 15.6 | 26.2 | 33.1 | download |
Method | Backbone | Schedule | AP | APS | APM | APL | Checkpoint |
---|---|---|---|---|---|---|---|
Mask R-CNN | R50-FPN | 1x | 33.8 | 12.0 | 31.5 | 51.8 | |
RefineMask | R50-FPN | 1x | 37.6 | 14.0 | 35.4 | 57.9 | download |
Method | AP | AP* | FPS |
---|---|---|---|
Mask R-CNN | 34.7 | 36.8 | 15.7 |
PointRend | 35.6 | 38.7 | 11.4 |
HTC | 37.4 | 40.7 | 4.4 |
RefineMask | 37.3 | 40.9 | 11.4 |
- Python 3.6+
- Pytorch 1.5.0
- mmcv-full 1.0.5
data
├── coco
| ├── annotations
│ │ │ ├── instances_train2017.json
│ │ │ ├── instances_val2017.json
│ │ │ ├── lvis_v0.5_val_cocofied.json
│ ├── train2017
│ │ ├── 000000004134.png
│ │ ├── 000000031817.png
│ │ ├── ......
│ ├── val2017
│ ├── test2017
├── lvis
| ├── annotations
│ │ │ ├── lvis_v1_train.json
│ │ │ ├── lvis_v1_val.json
│ ├── train2017
│ │ ├── 000000004134.png
│ │ ├── 000000031817.png
│ │ ├── ......
│ ├── val2017
│ ├── test2017
├── cityscapes
| ├── annotations
│ │ │ ├── instancesonly_filtered_gtFine_train.json
│ │ │ ├── instancesonly_filtered_gtFine_val.json
│ ├── leftImg8bit
│ | ├── train
│ │ ├── val
│ │ ├── test
Note: We used the lvis-v1.0 dataset which consists of 1203 categories.
./scripts/dist_train.sh ./configs/refinemask/coco/r50-refinemask-1x.py 8 work_dirs/r50-refinemask-1x
Note: The codes only support batch size 1 per GPU, and we trained all models with a total batch size 16x1. If you train models with a total batch size 8x1, the performance may drop. We will support batch size 2 or more per GPU later. You can use ./scripts/slurm_train.sh for training with multi-nodes. Multiple images per GPU during training has been supported now.
./scripts/dist_test.sh ./configs/refinemask/coco/r50-refinemask-1x.py 8 work_dirs/r50-refinemask-1x
@InProceedings{Zhang_2021_CVPR,
author = {Zhang, Gang and Lu, Xin and Tan, Jingru and Li, Jianmin and Zhang, Zhaoxiang and Li, Quanquan and Hu, Xiaolin},
title = {RefineMask: Towards High-Quality Instance Segmentation With Fine-Grained Features},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
month = {June},
year = {2021},
pages = {6861-6869}
}