Unsupervised domain adaptation algorithm design for robot perception

• Main Author: Tan, Jiawei
• Other Authors: Wang Dan Wei
• Format: Thesis-Master by Coursework
• Language: English
• Published: Nanyang Technological University 2021
• Subjects: Engineering::Electrical and electronic engineering::Control and instrumentation::Robotics; Engineering::Computer science and engineering::Computing methodologies::Image processing and computer vision
• Online Access: https://hdl.handle.net/10356/153407
• Institution: Nanyang Technological University

Depth completion task is an important direction in the current scene depth estimation field. The overall process of depth completion task is to restore the sparse depth map to a pixel-wide dense depth map with the aid of color image. Because the models that predict depth solely through monocular images is very sensitive to color and texture and has scale uncertainty, depth completion task is the most effective solution for realizing scene depth estimation. In this dissertation, depth completion task as a typical robot perception task is selected. We designed an end-to-end depth completion network: Attention based two-branch fusion network. The model is based on a dual-branch structure. And the two branches respectively focus more on color image input and depth information input for depth prediction. An attention based lightweight fusion module is proposed, which allocates spatial attention to the feature map from the two branches that need to be fused. This fusion module strengthens the transmission of important feature information and reduces the parameter scale of the model. Besides, a branch competition mechanism is designed to make the two branches strengthen their respective branch positions in the confrontation and achieve mutual optimization. Domain adaptation is to achieve consistent and excellent performance of the model in the feature space distribution of different data sets. We propose a style transfer based domain adaptation for depth completion. Based on this domain adaptation model, we have achieved high stability and high accuracy for our depth completion network in various weather environments. Our depth completion model and domain adaptation model for depth completion have been tested and evaluated in Virtual KITTI datasets with different weather environments. Evaluation shows that our model has desirable performance.