Detecting Small Objects in Thermal Images Using Single-Shot Detector

SSD (Single Shot Multibox Detector) is one of the most successful object detectors for its high accuracy and fast speed. However, features from shallow layer (mainly Conv4_3) lack semantic information, resulting in poor performance small objects. In this paper, we proposed DDSSD (Dilation Deconvolution Single Detector), an enhanced with a novel feature fusion module which can improve over detection. module, dilation convolution utilized to enlarge receptive field deconvolution adopted increase size maps layer. Our network achieves 79.7% mAP on PASCAL VOC2007 test 28.3% mmAP MS COCO test-dev at 41 FPS only 300x300 input using single Nvidia 1080 GPU. Especially, objects, 10.5% 22.8% FLIR thermal dataset, outperforming lot state-of-the-art detection algorithms both aspects