| Summary: | 碩士 === 國立交通大學 === 資訊科學與工程研究所 === 108 === Deep learning faces two barriers, having needs on large numbers of training data and a huge amount of numerical operations respectively. Distressingly, the process of labeling training data is strenuous. Complex deep learning networks typically spend time on the tens of thousands of computations, not to mention running multiple deep learning networks for different purposes simultaneously. Consequently, the automatic training data generation method, machine-train-machine migration learning, is proposed and verified in this work with the demonstration on Kinect. The design of the multi-task deep learning network is introduced between different networks to reduce computation time on the Kinect imitation. The performance of the multi-task deep learning network on skeleton detection and depth prediction, the performance of the software Kinect, are 93.5% in PCKh metric and 94.4% in A2 metric. Notably, the possible positive relationship of skeleton detection and depth prediction from the performance change on multi-task design is also mentioned. Most importantly, the time-complexity of the multi-task deep learning network is verified to indicate the actual detection efficiency improvement from the multi-task design with the processing speed of 52.14 fps. At last, the stereo pose information from our multi-task deep learning model is then utilized for the application on vision-based fitness e-coaching.
|
|---|
