Awesome

PS-MT

[CVPR'22] Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation

by Yuyuan Liu, Yu Tian, Yuanhong Chen, Fengbei Liu, Vasileios Belagiannis and Gustavo Carneiro

Computer Vision and Pattern Recognition Conference (CVPR), 2022

Installation

Please install the dependencies and dataset based on this installation document.

Getting start

Please follow this instruction document to reproduce our results.

Update

blender setting results in VOC12 dataset (under deeplabv3+ with resnet101)

Approach 1/16 (662) 1/8 (1323) 1/4 (2646) 1/2 (5291)
PS-MT (wandb_log) 78.79 80.29 80.66 80.87
- please note that, we update the blender splits list end with an extra 0 (e.g., 6620 for 662 labels) in the original directory.
- you can find the related launching scripts in here.
- In case you are using blender experiments (which are built on top of the high-quality labels), please compare with the results in this table.

Approach	1/16 (662)	1/8 (1323)	1/4 (2646)	1/2 (5291)
PS-MT (wandb_log)	78.79	80.29	80.66	80.87

Results

Pascal VOC12 dataset

augmented set

Backbone 1/16 (662) 1/8 (1323) 1/4 (2646) 1/2 (5291)
50 72.83 75.70 76.43 77.88
101 75.50 78.20 78.72 79.76
high quality set (based on res101)

1/16 (92) 1/8 (183) 1/4 (366) 1/2 (732) full (1464)
65.80 69.58 76.57 78.42 80.01

Backbone	1/16 (662)	1/8 (1323)	1/4 (2646)	1/2 (5291)
50	72.83	75.70	76.43	77.88
101	75.50	78.20	78.72	79.76

1/16 (92)	1/8 (183)	1/4 (366)	1/2 (732)	full (1464)
65.80	69.58	76.57	78.42	80.01

CityScape dataset

following the setting of CAC (720x720, CE supervised loss)

Backbone slid. eval 1/8 (372) 1/4 (744) 1/2 (1488)
50 ✗ 74.37 75.15 76.02
50 ✓ 75.76 76.92 77.64
101 ✓ 76.89 77.60 79.09
following the setting of CPS (800x800, OHEM supervised loss)

Backbone slid. eval 1/8 (372) 1/4 (744) 1/2 (1488)
50 ✓ 77.12 78.38 79.22

Backbone	slid. eval	1/8 (372)	1/4 (744)	1/2 (1488)
50	✗	74.37	75.15	76.02
50	✓	75.76	76.92	77.64
101	✓	76.89	77.60	79.09

Backbone	slid. eval	1/8 (372)	1/4 (744)	1/2 (1488)
50	✓	77.12	78.38	79.22

Training details

Some examples of training details, including:

VOC12 dataset in this wandb link.
Cityscapes dataset in this wandb link (w/ 1-teacher inference).

In details, after clicking the run (e.g., 1323_voc_rand1), you can checkout:

<img src="https://user-images.githubusercontent.com/102338056/167979073-1c1b3144-8a72-4d8d-9084-31d7fdab3e9b.png" width="26" height="22"> overall information (e.g., training command line, hardware information and training time).
<img src="https://user-images.githubusercontent.com/102338056/167978940-8c1f3d79-d062-4e7b-b56e-30b97d273ae8.png" width="26" height="22"> training details (e.g., loss curves, validation results and visualization)
<img src="https://user-images.githubusercontent.com/102338056/167979238-4847430f-aa0b-483d-b735-8a10b43293a1.png" width="26" height="22"> output logs (well, sometimes might crash ...)

Acknowledgement & Citation

The code is highly based on the CCT. Many thanks for their great work.

Please consider citing this project in your publications if it helps your research.

@article{liu2021perturbed,
  title={Perturbed and Strict Mean Teachers for Semi-supervised Semantic Segmentation},
  author={Liu, Yuyuan and Tian, Yu and Chen, Yuanhong and Liu, Fengbei and Belagiannis, Vasileios and Carneiro, Gustavo},
  journal={arXiv preprint arXiv:2111.12903},
  year={2021}
}

TODO

Code of deeplabv3+ for voc12
Code of deeplabv3+ for cityscapes