Compressing 3D-CNNs for Hardware-Efficient Object Classification in Video Streams

This paper presents a set of methods for effective compression of 3D-CNN deep learning architectures with a particular focus on embedded platforms. Deep learning models are massive regarding the number of parameters, especially in video processing. Such memory consumption poses a challenge when it comes to efficient deployment to embedded devices. The authors applied a series of quantization and pruning techniques and showed that it is possible to significantly reduce an 80M-parameters-large deep learning model with a negligible (approx. 1 perc.) decrease in accuracy.