Optimized Yolov3 Deployment on Jetson TX2 With Pruning and Quantization

Pruning and Quantization are commonly techniques deployed on deep convolutional networks to accelerate the model and reducing size, taking the resolution as a sacrifice. In embedded systems where computation power is limited due to power and cost constraints, pruning and quantization is mandatory for such networks especially where real-time processing is required, such as image classification in live video streams. In this paper, a pruned and quantized YOLOv3 model is deployed on Nvidia's industrial standard model Jetson TX2, which demonstrated an increase of 5 FPS in image classification via an 640×480 USB camera, while allocating only 16.7% storage on disk.