FFCA-Net: Stereo Image Compression via Fast Cascade Alignment of Side Information
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Deep stereo image compression, distributed source coding, stereo matching
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
Abstract: Multi-view compression technology, particularly stereo image compression (SIC), is vital for car cameras and 3D applications. Current SIC methods use a joint encoding and decoding compression structure, requiring both ends to use residual information at the same time. This creates a burden on the encoding terminal due to the need for camera cooperation. Interestingly, the distributed source coding theory suggests that efficient data compression of related sources can be achieved through independent encoding and joint decoding. This has led to the rapid development of deep-distributed SIC methods. However, these methods overlook the unique features of stereo-imaging tasks and cause high decoding latency. To overcome this, we suggest a Feature-based Fast Cascade Alignment network (FFCA-Net) to fully utilize the side information on the decoder. FFCA uses a step-by-step alignment approach. Initially, FFCA uses a feature domain patch-matching module based on stereo priors to reduce redundancy and noise. Then, we use an hourglass-based sparse stereo refinement network to further align inter-image features at a lower computational cost. Additionally, we've created a lightweight, high-performance feature fusion network, the Fast Feature Fusion Network (FFF), to decode the aligned features. Tests on InStereo2K, KITTI, and Cityscapes datasets show our method is significantly better than traditional and learning-based SIC methods. Specifically, our method achieves 3 to 10 times faster decoding speed than other methods.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 4896
Loading