Are you sure you want to delete this task? Once this task is deleted, it cannot be recovered.
slfan 6b0c8e031f | 2 years ago | |
---|---|---|
README.md | 2 years ago | |
arc.png | 2 years ago | |
qual.jpg | 2 years ago | |
quan.jpg | 2 years ago |
This is the official implementaion of GANet paper "Quality-Aware Optimization for RGB-D Salient Object Detection".
Despite the tremendous progress in RGB-D salient object detection (SOD), existing methods often regard this task as a per-pixel classification problem optimized by the binary cross-entropy loss, then utilize several other metrics to evaluate the segmentation predictions from different aspects (e.g., structural or global properties). The disagreement between the optimization and evaluation goals could result in a suboptimum and is an underexplored issue in many segmentation tasks. To unify the metric for both optimization and evaluation, we model this problem as an image quality assessment task. Concretely, our solution consists of a generator subnetwork and a discriminator subnetwork, where the former generates segmentation predictions according to specific vision tasks, while the latter serves as an assistant and provides a correct evaluation goal to optimize the former. For distinguishing from the traditional generative adversarial networks, we name it Generative Assistant Network (GANet) as its discriminator supports being independently trained till convergence in a self-supervised way. Thanks to the generator and discriminator of the proposed framework being orthogonal to specific network design, we can apply it to many segmentation tasks. Starting with RGB-D SOD, we verify its effectiveness on various segmentation tasks, i.e., camouflaged object detection and polyp segmentation, and always observe a significant performance improvement. Moreover, to further promote the accuracy of RGB-D SOD, we design an effective saliency generator by imitating the generation of biological saliency. Extensive experiments indicate that the proposed method can obtain a new state-of-the-art performance ( ~25\ mean absolute error drop).
Quantitative comparison
Qualitative comparison
Results of our JL-DCF model on 7 benchmark datasets (NJUD, NLPR, STERE, DES, LFSD, SIP, DUT) can be found below:
link: https://pan.baidu.com/s/1zs0H0uy14NCU2sSwrYmeUA with code: kw2d
Please drop me an email for further problems or discussion: slfan@pku.edu.cn
Dear OpenI User
Thank you for your continuous support to the Openl Qizhi Community AI Collaboration Platform. In order to protect your usage rights and ensure network security, we updated the Openl Qizhi Community AI Collaboration Platform Usage Agreement in January 2024. The updated agreement specifies that users are prohibited from using intranet penetration tools. After you click "Agree and continue", you can continue to use our services. Thank you for your cooperation and understanding.
For more agreement content, please refer to the《Openl Qizhi Community AI Collaboration Platform Usage Agreement》