Improved Single Shot Detector with Enhanced Hard Negative Mining Approach

Image classification and tiny-object detection are challenging tasks in computer vision domain. This is primarily due to their ability to tackle real-world problems, such as de-veloping self-driving cars, robot navigation, surveillance systems, and monitoring road safety. A hard negative mining approach is predominately utilized to train object detection networks, which use positive and negative samples to increase network gains, but the selection of negative samples is an expensive process as the network identifies many duplicates. Numerous research findings are being carried out to enhance hard negative mining. This research addresses the drawbacks of existing techniques in the hard negative mining approach and proposes utilizing medium priors to improve network performance. Medium priors can be defined as anchor boxes with 20 % to 50 % overlap with ground truth boxes. Since the tiny objects are much smaller than other objects in a frame, considering medium priors significantly enhances the detection probability. The proposed metric has been evaluated using Single Shot Multibox Detector (SSD) architec-ture. Experimental results on PASCAL VOC datasets indicate that the average precision of tiny objects such as potted plants and bottles increased by 4 % and 3.9 % with an overall increase in mAP of 0.9%.