Towards Accurate One-Stage Object Detection with AP-Loss


Chen, Kean and Li, Jianguo and See, John Su Yang and Ji, Wang and Duan, Lingyu and Chen, Zhibo and He, Changwei and Zou, Junni (2019) Towards Accurate One-Stage Object Detection with AP-Loss. In: IEEE International Conference on Computer Vision and Pattern Recognition (CVPR), 16-20 June 2019, Long Beach Convention Center, Long Beach, CA, United States.

[img] Text
79.pdf - Published Version
Restricted to Repository staff only

Download (1MB)


One-stage object detectors are trained by optimizing classification-loss and localization-loss simultaneously, with the former suffering much from extreme foreground-background class imbalance issue due to the large number of anchors. This paper alleviates this issue by proposing a novel framework to replace the classification task in one-stage detectors with a ranking task, and adopting the Average-Precision loss (AP-loss) for the ranking problem. Due to its non-differentiability and non-convexity, the AP-loss cannot be optimized directly. For this purpose, we develop a novel optimization algorithm, which seamlessly combines the error-driven update scheme in perceptron learning and backpropagation algorithm in deep networks. We verify good convergence property of the proposed algorithm theoretically and empirically. Experimental results demonstrate notable performance improvement in state-of-the-art one-stage detectors based on AP-loss over different kinds of classification-losses on various benchmarks, without changing the network architectures.

Item Type: Conference or Workshop Item (Paper)
Uncontrolled Keywords: Object Detection, Object monitors (Computer software)
Subjects: Q Science > QA Mathematics > QA71-90 Instruments and machines > QA75.5-76.95 Electronic computers. Computer science > QA76.75-76.765 Computer software
Divisions: Faculty of Computing and Informatics (FCI)
Depositing User: Ms Rosnani Abd Wahab
Date Deposited: 08 Oct 2021 03:39
Last Modified: 08 Oct 2021 03:39


Downloads per month over past year

View ItemEdit (login required)