Dissected 3D CNNs: Temporal skip connections for efficient online video processing

Convolutional Neural Networks with 3D kernels (3D CNNs) currently achieve state-of-the-art results in video recognition tasks due to their supremacy extracting spatiotemporal features within frames. There have been many successful CNN architectures surpassing successively. However, nearly all of them are designed operate offline, creating several serious handicaps during online operation. Firstly, conventional CNNs not dynamic since output represent the complete input clip instead most recent frame clip. Secondly, they temporal resolution-preserving inherent downsampling. Lastly, constrained be used fixed size limiting flexibility. In order address these drawbacks, we propose dissected CNNs, where intermediate volumes network and propagated over depth (time) dimension for future calculations, substantially reducing number computations at For action classification, version ResNet models performs 77%–90% fewer operation while achieving ∼5% better classification accuracy on Kinetics-600 dataset than 3D-ResNet models. Moreover, advantages demonstrated by deploying our approach onto vision tasks, which consistently improved performance.