A Local Top-Down Module for Object Detection with Multi-scale Features

Shihua Huang, Lu Wang, Peiyu Yang, Qingxu Deng

Abstract: Object detection methods based on deep models and multi-scale features have achieved the state-of-the-art performance. However, since each feature layer operates independently, several issues such as box-in-box detections and less effective performance on small objects need to be addressed. In this paper, we tackle these issues by integrating contextual and semantic information from higher layer features into the prediction layer. Existing methods adopting similar ideas mostly apply full top-down modules, which may increase computational loads significantly. Instead, we present an efficient while general local top-down module, in which each prediction layer is integrated only with the upsampled features from its two succeeding layers. Experimental results show that the proposed algorithm performs favorably against the state-of-the-art methods on the VOC, COCO and HollywoodHeads datasets, while introducing little computational overhead. Compared with methods using full top-down modules, the proposed algorithm achieves comparable or higher accuracy while operates at a higher frame rate. The code is available at https://github.com/Hshihua/Local-Top-Down-Detection-Network .