Efficient YOLO: A Lightweight Model for Embedded Deep Learning Object Detection

It is essential to pursue efficiency for on-road object detection task. To incorporate deep model into embedded devices while maintaining high accuracy, in this paper, an Efficient YOLO framework is rebuilt based on traditional YOLOv3. Firstly, an iterative initialization strategy is designed to ensure the network sparsity in the initial training. Then comprehensive pruning schemes including layer-level and channel-wise pruning are proposed to lighten the model parameters.With the support of external dataset, the detection accuracy remains at a high level. Compared with the orignal version, our model shrinks the model size by 96.93% and calculation amount by 84.36%. The inference speed is improved 2.23 times on NVIDIA Jetson TX2 platform. Finally, we achieve a mAP of 0.492 on the testing dataset, and rank the top accuracy of ICME competition.