BMaskR-CNN
This code is developed on Detectron2
Boundary-preserving Mask R-CNN
ECCV 2020
Tianheng Cheng, Xinggang Wang, Lichao Huang, Wenyu Liu
Video from Cam看世界 on Youtube.
Abstract
Tremendous efforts have been made to improve mask localization accuracy in instance segmentation. Modern instance segmentation methods relying on fully convolutional networks perform pixel-wise classification, which ignores object boundaries and shapes, leading coarse and indistinct mask prediction results and imprecise localization. To remedy these problems, we propose a conceptually simple yet effective Boundary-preserving Mask R-CNN (BMask R-CNN) to leverage object boundary information to improve mask localization accuracy. BMask R-CNN contains a boundary-preserving mask head in which object boundary and mask are mutually learned via feature fusion blocks. As a result,the mask prediction results are better aligned with object boundaries. Without bells and whistles, BMask R-CNN outperforms Mask R-CNN by a considerable margin on the COCO dataset; in the Cityscapes dataset,there are more accurate boundary groundtruths available, so that BMaskR-CNN obtains remarkable improvements over Mask R-CNN. Besides, it is not surprising to observe that BMask R-CNN obtains more obvious improvement when the evaluation criterion requires better localization (e.g., AP75)
Models
COCO
| Method | Backbone | lr sched | AP | AP50 | AP75 | APs | APm | APl | download |
|---|---|---|---|---|---|---|---|---|---|
| Mask R-CNN | R50-FPN | 1x | 35.2 | 56.3 | 37.5 | 17.2 | 37.7 | 50.3 | - |
| PointRend | R50-FPN | 1x | 36.2 | 56.6 | 38.6 | 17.1 | 38.8 | 52.5 | - |
| BMask R-CNN | R50-FPN | 1x | 36.6 | 56.7 | 39.4 | 17.3 | 38.8 | 53.8 | model |
| BMask R-CNN | R101-FPN | 1x | 38.0 | 58.6 | 40.9 | 17.6 | 40.6 | 56.8 | model |
| Cascade Mask R-CNN | R50-FPN | 1x | 36.4 | 56.9 | 39.2 | 17.5 | 38.7 | 52.5 | - |
| Cascade BMask R-CNN | R50-FPN | 1x | 37.5 | 57.3 | 40.7 | 17.5 | 39.8 | 55.1 | model |
| Cascade BMask R-CNN | R101-FPN | 1x | 39.1 | 59.2 | 42.4 | 18.6 | 42.2 | 57.4 | model |
Cityscapes
- Initialized from ImagetNet pre-training.
| Method | Backbone | lr sched | AP | download |
|---|---|---|---|---|
| PointRend | R50-FPN | 1x | 35.9 | - |
| BMask R-CNN | R50-FPN | 1x | 36.2 | model |
Results
Left: AP curves of Mask R-CNN and BMask R-CNN under different mask IoU thresholds on the COCO val2017 set, the improvement becomes more significant when IoU increases. Right: Visualizations of Mask R-CNN and BMask R-CNN. BMask R-CNN can output more precise boundaries and accurate masks than Mask R-CNN.
Usage
Install Detectron2 following the official instructions
Training
specify a config file and train a model with 4 GPUs
cd projects/BMaskR-CNN
python train_net.py --config-file configs/bmask_rcnn_R_50_FPN_1x.yaml --num-gpus 4
Evaluation
specify a config file and test with trained model
cd projects/BMaskR-CNN
python train_net.py --config-file configs/bmask_rcnn_R_50_FPN_1x.yaml --num-gpus 4 --eval-only MODEL.WEIGHTS /path/to/model
Citation
@article{ChengWHL20,
title={Boundary-preserving Mask R-CNN},
author={Tianheng Cheng and Xinggang Wang and Lichao Huang and Wenyu Liu},
booktitle={ECCV},
year={2020}
}
57 Nov 27, 2022
13 Jan 18, 2022
4 Apr 5, 2022
25.5k Jan 7, 2023
1.2k Jan 4, 2023
2 Dec 4, 2021
22.5k Jan 4, 2023
5 Nov 3, 2022
221 Dec 30, 2022
79 Oct 8, 2022
21 Nov 20, 2022
37 Dec 4, 2022
54 Dec 15, 2022
98 Dec 7, 2022
32 Dec 13, 2022
42 Nov 24, 2022
44 Dec 14, 2022
242 Dec 20, 2022
207 Jan 5, 2023