GCoNet
The official repo of the CVPR 2021 paper Group Collaborative Learning for Co-Salient Object Detection .
Trained model
Download final_gconet.pth (Google Drive). And it is the training log.
Put final_gconet.pth at GCoNet/tmp/GCoNet_run1.
Run test.sh for evaluation.
Data Format
Put the DUTS_class (training dataset from GICD), CoCA, CoSOD3k and Cosal2015 datasets to GCoNet/data as the following structure:
GCoNet
├── other codes
├── ...
│
└── data
├──── images
| ├── DUTS_class (DUTS_class's image files)
| ├── CoCA (CoCA's image files)
| ├── CoSOD3k (CoSOD3k's image files)
│ └── Cosal2015 (Cosal2015's image files)
│
└────── gts
├── DUTS_class (DUTS_class's Groundtruth files)
├── CoCA (CoCA's Groundtruth files)
├── CoSOD3k (CoSOD3k's Groundtruth files)
└── Cosal2015 (Cosal2015's Groundtruth files)
Usage
Run sh all.sh for training (train_GPU0.sh) and testing (test.sh).
Prediction results
The co-saliency maps of GCoNet can be found at Google Drive.
Note and Discussion
In your training, you can usually obtain slightly worse performance on CoCA dataset and slightly better perofmance on Cosal2015 and CoSOD3k datasets. The performance fluctuation is around 1.0 point for Cosal2015 and CoSOD3k datasets and around 2.0 points for CoCA dataset.
We observed that the results on CoCA dataset are unstable when train the model multiple times, and the performance fluctuation can reach around 1.5 ponits (But our performance are still much better than other methods in the worst case).
Therefore, we provide our used training pairs and sequences with deterministic data augmentation to help you to reproduce our results on CoCA. (In different machines, these inputs and data augmentation are different but deterministic.) However, there is still randomness in the training stage, and you can obtain different performance on CoCA.
There are three possible reasons:
- It may be caused by the challenging images of CoCA dataset where the target objects are relative small and there are many non-target objects in a complex environment.
- The imperfect training dataset. We use the training dataset in GICD, whose labels are produced by the classification model. There are some noisy labels in the training dataset.
- The randomness of training groups. In our training, two groups are randomly picked for training. Different collaborative training groups have different training difficulty.
Possible research directions for performance stability:
- Reduce label noise. If you want to use the training dataset in GICD to train your model. It is better to use multiple powerful classification models (ensemble) to obtain better class labels.
- Deterministic training groups. For two collaborative image groups, you can explore different ways to pick the suitable groups, e.g., pick two most similar groups for hard example mining.
It is a potential research direction to obtain stable results on such challenging real-world images. We follow other CoSOD methods to report the best performance of our model. You need to train the model multiple times to obtain the best result on CoCA dataset. If you want more discussion about it, you can contact me ([email protected]).
Citation
@inproceedings{fan2021gconet,
title={Group Collaborative Learning for Co-Salient Object Detection},
author={Fan, Qi and Fan, Deng-Ping and Fu, Huazhu and Tang, Chi-Keung and Shao, Ling and Tai, Yu-Wing},
booktitle={CVPR},
year={2021}
}
Acknowledgements
Zhao Zhang gives us lots of helps! Our framework is built on his GICD.
4 Aug 25, 2022
37 Dec 8, 2022
28 Oct 20, 2022
33 Jan 5, 2023
6.5k Jan 9, 2023
334 Dec 23, 2022
67 Dec 4, 2022
29 Dec 10, 2022
146 Dec 24, 2022
76 Jan 2, 2023
184 Dec 11, 2022
41 Oct 21, 2022
98 Nov 16, 2022
305 Dec 19, 2022
128 Dec 8, 2022
149 Dec 22, 2022
269 Dec 21, 2022