Learning in Compressed Domain via Knowledge Transfer

Hanyue Tu, Li Li, Wengang Zhou, Houqiang Li

Keywords: compressed-domain vision, image compression, knowledge transfer

Abstract: Learning in compressed domain aims to perform vision tasks directly on compressed latent representations instead of reconstructed images. Existing reports show that learning in compressed domain can achieve a comparable performance compared to that in pixel domain for certain compression models. However, we observe that when using the state-of-the-art learned compression models, the performance gap between compressed-domain and pixel-domain vision tasks is still large due to the lack of some natural inductive biases in pixel-domain convolutional neural networks. In this paper, we attempt to address this problem by transferring knowledge from pixel domain to compressed domain. We first modify neural networks for pixel-domain vision tasks to better suit compressed-domain inputs. In addition, we propose a knowledge transfer loss to narrow the gap between compressed domain and pixel domain. Experimental results on classification and instance segmentation show that the proposed method improves the accuracy of compressed-domain vision tasks significantly, which even outperforms learning on reconstructed images while avoiding the computational cost for image reconstruction.

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.

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

Submission Guidelines: Yes

Please Choose The Closest Area That Your Submission Falls Into: Deep Learning and representational learning

TL;DR: We propose learning in compressed domain by transferring the knowledge learned in pixel domain.