Rep-Tracker: a lightweight tiny ball tracking method using structure re-parameterization

TL;DR: In this paper , a lightweight tiny ball tracking method is developed based on deep learning call Rep-Tracker, which incorporates the structure re-parameterization strategy on the basis of the pruned VGG-style model to construct a tennis tracking network.

Abstract: Abstract Vision-based object tracking techniques are used to analyze sport competition videos. The latest tennis tracking models achieve very high accuracy. However, high precision networks often imply higher computational complexity, which in turn brings higher hardware requirements. Also, part of the networks suffer from the problem of different degrees of adaptation to different courses. In this paper, a lightweight tiny ball tracking method is developed based on deep learning call Rep-Tracker, which incorporates the structure re-parameterization strategy on the basis of the pruned VGG-style model to construct a tennis tracking network. Firstly, the video frames will be sent into a pruned VGG-16 encoder follow by a decoder composed of deconvolution, and heatmap will be output after a softmax layer. Based on the encoder-decoder structure , pruning methods are used to reduce the number of non-essential convolutional layers and channels in the network. Secondly, residual connections are added at each layers to form a multi-branch block at training time. By using a structural re-parameterization strategy, each branch in a multi-branch block will be equivalently transformed into a shape consistent convolution. The trained blocks are convert into a single convolution layer for inference. Finally, based on the heatmap, the position of the target object is obtained by Gaussian distribution function 1 Springer Nature 2021 L A T E X template Article Title and Hough gradient function. In the tennis tracking task our method achieves a high accuracy with GFLOPs only 1.29 on RTX 3080 GPU.