| Abstrakt: |
Current one-stage object detection methods use dense prediction to generate classification and regression results at the same point on the feature map. Due to the different task attributes, classification and regression are typically trained using separate detection heads, which may result in different feature areas being focused on. However, they ultimately act on the same object, especially in the post-processing stage, where we hope they have the same performance. This inherent contradiction can seriously affect the performance of the detector. To solve this problem, we propose a flexible and effective decouple and align classification and regression one-stage object detector (DAOD), based on different aspects to decouple and align the two subtasks. Specifically, we first propose a regression subtask spatial decouple module to solve the regression spatial sensitivity problem by efficiently sampling the information of the regression result map to strengthen localization. Then, we propose a dynamic aligned label assignment strategy for sample selection, guiding the network to focus on more aligned features during training. Finally, we introduce harmonic supervision to align results while ensuring the independence of the respective task. With the negligible additional overhead, extensive experiments on the COCO dataset demonstrate the effectiveness of our DAOD. Notably, DAOD with ResNeXt-101-64 × 4d-DCN backbone achieves 50.0 AP at single-model single-scale testing on MS-COCO test-dev. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink