Data Driven Sensing for Action Recognition Using Deep Convolutional Neural Networks

摘要

Tasks such as action recognition requires high quality features for accurate inference. But the use of high resolution and large volume of video data poses a significant challenge for inference in terms of storage and computational complexity. In addition, compressive sensing as a potential solution to the aforementioned problems has been shown to recover signals at higher compression ratios with loss in information. Hence, a framework is required that performs good quality action recognition on compressively sensed data. In this paper, we present data-driven sensing for spatial multiplexers trained with combined mean square error (MSE) and perceptual loss using Deep convolutional neural networks. We employ subpixel convolutional layers with the 2D Convolutional Encoder-Decoder model, that learns the downscaling filters to bring the input from higher dimension to lower dimension in encoder and learns the reverse, i.e. upscaling filters in the decoder. We stack this Encoder with Inflated 3D ConvNet and train the cascaded network with cross-entropy loss for Action recognition. After encoding data and undersampling it by over 100 times (10 × 10) from the input size, we obtain 75.05% accuracy on UCF-101 and 50.39% accuracy on HMDB-51 with our proposed architecture setting the baseline for reconstruction free action recognition with data-driven sensing using deep learning. We experimentally infer that the encoded information from such spatial multiplexers can directly be used for action recognition.