YOLO-compact: An Efficient YOLO Network for Single Category Real-time Object Detection
Yonghui Lu, Langwen Zhang, Wei Xie
Abstract
In practical applications, the number of category in object detection is always single. In this paper, an efficient YOLO-compact network designed for single category real-time object detection is proposed. This paper first explored a series of methods for converting a large and deep network to a compact and efficient network, through a series of ablation experiments. Then these methods were assembled to YOLOv3 network, which obtains the YOLO-compact's infrastructure. A network structure design approach that separates the down sampling layer from all network modules was proposed, which facilitates the modular design of the network. The shortcut structure in the RFB module is changed to the direct connection structure, and the 1×3+3×1+3×3 convolution structure is used instead of 5×5 convolution, which obtains efficient RFB-c module. The residual bottleneck block has been improved, by removing the last 1×1 convolution layer and using 3×3 depth wise separable convolution. Since pedestrian is the most representative object in practical applications, this paper uses the result on person category in VOC2007 test set to represent the network performance. The model size of YOLO-compact is only 9MB, which is 3.7 times smaller than tiny-yolov3, 6.7 times smaller than tiny-yolov2, and 26 times smaller than YOLOv3. The AP result of YOLO-compact is 86.85%, which is 37% higher than tiny-yolov3, 32% higher than tiny-yolov2, and even 2.7% higher than the YOLOv3.