Convolutional 3D in Activity Recognition -A Review

Activity recognition in videos using deep learning has shown phenomenal progress in the last decade. The community of computer vision has been working on video data for about a decade and solved many uncertainties. Various research groups have presented many convolutional neural network architectures to solve the issues related with classification, and many more computer vision tasks. All these networks were about two dimensional image data. Recently research community of Face book introduced a network architecture called C3D network which have three dimensional convolution layers. The C3D network is performing well in activity recognition from large-scale videos, video classification tasks. This article is focused on importance of C3D network by mentioning the drawbacks of hand crafted as well as deep learning methods, architecture of C3D network, contrasts in 2D and 3D CNNs, review on different deep learning models employed for activity detection from videos and compared the performance of various anomaly detection approaches with the proposed C3D network.