Depth estimation from infrared video using local-feature-flow neural network

Depth estimation is essential for infrared video processing. In this paper, a novel depth estimation method, called local-feature-flow neural network (LFFNN), is proposed for generating depth maps for each frame of an infrared video. LFFNN extracts local features of a frame with the addition of inter-frame features, which is extracted from the previous frames on the corresponding region in the infrared video. LFFNN is designed for extracting the local features flow in the infrared video, learning better depth-related features through three control gates by inter-frame features propagation as the video progresses. After feature extraction, a pixel-level classifier is created to estimate depth level of different pixels in the infrared video. Our proposed approach achieves state-of-the-art depth estimation performances on the test dataset.