Learning Richer Features in Deep CNN for Object Detection

Recently developed object detectors employ a Deep Convolutional Neural Network (DCNN) by adding the number of feature layers with a pyramidal shape. Owing the occurrence of Feature Pyramid Network (FPN), the representation of the DCNN's ability was largely improved. Although these object detectors with feature pyramid achieve encouraging results, they have some limitations that didn't fully explore richer information in the deep layers, which is essential for the Object Detection task. In this work, we propose Multi-Level Spatial Pyramid Pooling module to construct more effective feature network for detecting objects of different scales. It is an integration of three pooling schemes of different kernel sizes. Besides, we design novel backbone based Darknet as our feature extractor. First, we fuse multi-level features extracted by new backbone followed by FPN style as the basic features. Then we feed the fused feature to the module to generate richer semantic feature maps at different scales. Finally, we gather up the richer feature maps in the former step for the detectors. To evaluate the effectiveness of the proposed method, experiments are conducted on two major benchmarks, which is PASCAL VOC 2007 and MS COCO dataset, and demonstrates that the ML-SPP module achieve comparable results with high efficiency.