A Simple Framework for Low-Resolution Detection with High-resolution Knowledge

Jiahui Chen

Keywords: object detection, low-resolution, knowledge distillation

TL;DR: We propose a simple distillation-based framework to improve detection on low-resolution images.

Abstract: This paper dedicates to improving object detection performance on low-resolution images. The intuitive way is to distill the high-resolution knowledge from models trained over high-resolution images, shorted as cross-resolution distillation. Unfortunately, most of existing conventional distillation methods focus on the knowledge distillation with same-resolution images in both teacher and student. Directly applying these methods for the cross-resolution distillation results in limited improvement. To address this issue, we introduce a simple yet effective framework, i.e., LRDet. The key in LRDet is the bridge branch, acting as an intermediate status between teacher and student. With the bridge branch, LRDet can i) align the resolution and supervision targets between the high-resolution teacher and the low-resolution student, and ii) then transfer the high-resolution knowledge smoothly and effectively. Experiments demonstrate that LRDet consistently improves various well-known detectors on low-resolution images, e.g., from 35.4 mAP to 37.8 mAP with RetinaNet-R50 on MS COCO using 600 × 1000 input. Meanwhile, it is easy to utilize large teachers in LRDet as the conventional distillation methods do, which can further improve the low-resolution performance. For example, RetinaNet-R50 with 600 × 1000 resolution can achieve 39.7 mAP when distilling from RetinaNet-X101.

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: Applications (eg, speech processing, computer vision, NLP)