Trident SSD: A Trident Single-Shot Multibox Object Detector with Deconvolution

Abstract

Scale variation is one of the key challenges in the object detection area, which limits the precision of detection methods like Single shot multibox detector (SSD). This paper proposes a detection method based on SSD, which focuses on handling scale variation and better detection performance of small objects. Our method, called TridentSSD, have an architecture with three branches, which are respectively responsible for detecting different scales of objects, solving the problem of scale variation while training. Then we augment small object branch with deconvolution module and feature fusion methods to improve precision, especially for small object detection. During training, we modify the original matching rules to generate training samples. Consequently, we can use objects of different scales to train the corresponding branches respectively. Finally, Experiments have done on both PASCAL VOC2007 and VOC2012 datasets. Results show that, with an input size of 300×300, our TridentSSD achieves better performance compared to the benchmark method SSD.

Sewing gesture image detection method based on improved SSD model

“…However, the relatively large depth of the model structure drives down its detection speed. Chen et al [6] amplified and extracted small target feature regions from SSD networks, extracted semantic information from multiple deep feature layers, and improved the medium-scale feature representation of mesoscale objects. However, their method also has low detection speed.…”