Improve Real-time Object Detection with Feature Enhancement

Real-time object detection is a critical task in the field of computer vision. Current methods suffers from pooling errors and classification imbalance. In this paper, we proposed a feature enhanced framework to improve the feature extraction for efficient real-time object detection. Particularly, the framework is compromised with five optimizations on feature enhancement: bilinear interpolation up-sampling, network fully sharing, single score map, upward rounding quantization and single layer output. We performed comprehensive experiments on a publicly available dataset DACSDC, and the achieved mean average precision (mAP) is 0.6317 with a throughput of 24.67 frame per second (FPS). Our work ranked the 6th out of 24 teams in the 2018 system design contest on the 55th Design Automatic Conference (DAC).