Authors: Ge-Peng Ji, Deng-Ping Fan, Yu-Cheng Chou, Dengxin Dai, Alexander Liniger, & Luc Van Gool.
This official repository contains the source code, prediction results, and evaluation toolbox of Deep Gradient Network (accepted by Machine Intelligence Research 2023), also called DGNet. The technical report can be found at arXiv. The following is a quick video to introduce our work:
MIR.Series.Deep.Gradient.Learning.for.Efficient.Camouflaged.Object.Detection-2.mp4
Figure 1: We present the scatter relationship between the performance weighted F-measure and parameters of all competitors on the CAMO-Test. These scatters are in various colors for better visual recognition and are also corresponding to the histogram (Right).
The larger the size of the colored scatter point, the heavier the model parameter. (Right) We also report the parallel histogram comparison of the model's parameters, MACs, and performance.
-
Novel supervision. We propose to excavate the texture information via learning the object level gradient rather than using boundary-supervised or uncertainty-aware modeling.
-
Simple but efficient. We decouple all the heavy designs as much as we can, yielding a simple but efficient framework. We hope this framework could serve as a baseline learning paradigm for the COD field.
-
Best trade-off. Our vision is to achieve a new SOTA with the best performance-efficiency trade-off on existing cutting-edge COD benchmarks.
- [2023/08/01] All onedrive download links are expired due to unknown technical problems. Now all download links are replaced with Google Drive links.
- [2022/11/18] The segmentation results on CHAMELEON dataset are available at GoogleDrive: DGNet (CAHMELEON) and DGNet-S (CHEMELEON).
- [2022/11/14] We convert the PyTorch model to ONNX model that is easier to access hardware optimizations and Huawei-OM model that supports Huawei Ascend series AI processors. More details can be found at lib_ascend.
- [2022/11/03] We add the support for the PVTv2 Transformer backbone, achieving excited performance again on COD benchmarks. Please enjoy it -> (link)
- [2022/08/06] Our paper has been accepted by Machine Intelligence Research (MIR).
- [2022/05/30] 🔥 We release the implementation of DGNet with different AI frameworks: Pytorch-based and Jittor-based.
- [2022/05/30] Thank @Katsuya Hyodo for adding our model into PINTO. This is a repository for storing models that have been inter-converted between various frameworks (e.g., TensorFlow, PyTorch, ONNX).
- [2022/05/25] Releasing the codebase of DGNet (Pytorch) and whole COD benchmarking results (20 models).
- [2022/05/23] Creating repository.
This project is still a work in progress, and we invite all to contribute to making it more accessible and useful. If you have any questions about our paper, feel free to contact me via e-mail ([email protected] & [email protected] & [email protected]). If you are using our code and evaluation toolbox for your research, please cite this paper (BibTeX).
Figure 2: Overall pipeline of the proposed DGNet, It consists of two connected learning branches, i.e., context encoder and texture encoder.
Then, we introduce a gradient-induced transition (GIT) to collaboratively aggregate the feature that is derived from the above two encoders. Finally, a neighbor-connected decoder (NCD [1]) is adopted to generate the prediction.
Figure 3: Illustration of the proposed gradient-induced transition (GIT).
It uses a soft grouping strategy to provide parallel nonlinear projections at multiple fine-grained sub-spaces, which enables the network to probe multi-source representations jointly.
References of neighbor-connected decoder (NCD) benchmark works [1] Concealed Object Detection. TPAMI, 2023.
We provide various versions for different deep learning platforms, including PyTorch and Jittor libraries. Note that we only report the results of the Pytorch-based DGNet in our manuscript.
-
For the Pytorch users, please refer to our lib_pytorch.
-
For the Jittor users, please refer to our lib_jittor.
-
For the Huawei Ascend users, please refer to our lib_ascend.
The whole benchmark results can be found at Google Drive. Here are quantitative performance comparisons from three perspectives. Note that we used the Matlab-based toolbox to generate the reported metrics.
Figure 4: Quantitative results in terms of full metrics for cutting-edge competitors, including 8 SOD-related and 12 COD-related, on three test datasets: NC4K-Test, CAMO-Test, and COD10K-Test. @R means the ranking of the current metric, and Mean@R indicates the mean ranking of all metrics.
Figure 5: Super-classes (i.e., Amphibian, Aquatic, Flying, Terrestrial, and Other) on the COD10K-Test of the proposed methods (DGNet & DGNet-S) and other 20 competitors. Symbol \uparrow indicates the higher the score, the better, and symbol \downarrow indicates the lower, the better. The best score is marked in bold.
Figure 6: Sub-class results on COD10K-Test of 12 COD-related and 8 SOD-related baselines in terms of structure measure (\mathcal{S}_\alpha), where Am., Aq., Fl., Te., and Ot. represent Amphibian, Aquatic, Flying, Terrestrial, and Other, respectively. CDL., GP.Fish, and LS.Dragon denote Crocodile, and GhostPipeFish, LeafySeaDragon, respectively. The best score is marked in bold.
Please cite our paper if you find the work useful:
@article{ji2023gradient,
title={Deep Gradient Learning for Efficient Camouflaged Object Detection},
author={Ji, Ge-Peng and Fan, Deng-Ping and Chou, Yu-Cheng and Dai, Dengxin and Liniger, Alexander and Van Gool, Luc},
journal={Machine Intelligence Research},
pages={92-108},
volume={20},
issue={1},
year={2023}
}