Multimodal Egocentric Activity Recognition Using Multi-stream CNN

Abstract

Egocentric activity recognition (EAR) is an emerging area in the field of computer vision research. Motivated by the current success of Convolutional Neural Network (CNN), we propose a multi-stream CNN for multimodal egocentric activity recognition using visual (RGB videos) and sensor stream (accelerometer, gyroscope, etc.). In order to effectively capture the spatio-temporal information contained in RGB videos, two types of modalities are extracted from visual data: Approximate Dynamic Image (ADI) and Stacked Difference Image (SDI). These image-based representations are generated both at clip level as well as entire video level, and are then utilized to finetune a pretrained 2D-CNN called MobileNet, which is specifically designed for mobile vision applications. Similarly for sensor data, each training sample is divided into three segments, and a deep 1D-CNN network is trained (corresponding to each type of sensor stream) from scratch. During testing, the softmax scores of all the streams (visual + sensor) are combined by late fusion. The experiments performed on multimodal egocentric activity dataset demonstrates that our proposed approach can achieve state-of-the-art results, outperforming the current best handcrafted and deep learning based techniques.