MDSSD: Multi-scale Deconvolutional Single Shot Detector for Small Objects

For most of the object detectors based on multi-scale feature maps, the shallow layers are mainly responsible for small object detection due to their fine details. However, the performance of detecting small object instances is still less satisfactory because of the deficiency of semantic information on shallow features. For the top semantic features, the representation of fine details for small objects are potentially wiped out. In this paper, we design a Multi-scale Deconvolutional Single Shot Detector (MDSSD), especially for the detection of small objects. In MDSSD, to generate features with strong representational power for small object instances, we add the high-level features with rich semantic information to the low-level features via deconvolution Fusion Block. It is noteworthy that multiple high-level features with different scales are upsampled simultaneously in our framework. Afterwards, we implement the skip connections to form more descriptive feature maps for small objects and predictions are made on these new fusion features. Our proposed framework achieves 78.6% mAP on PASCAL VOC2007 test and 26.8% mAP on MS COCO test-dev2015 at 38.5 FPS with only 300*300 input. The results outperform baseline SSD by 1.1 and 1.7 points respectively, especially with 2 -- 5 points improvement on some small objects categories.