A Practical Stereo Depth System for Smart Glasses

Figure 1. (a) Overview. The user captures a stereo pair with smart glasses. The images are sent to the user's phone for processing, including online rectification and fast disparity prediction via neural networks. The predicted disparity map is then used to generate visual effects by rendering novel views. (b) Zero-shot accuracy. Our stereo network, Argos, achieves high accuracy despite not being trained on these datasets, being quantized to 8 bits, and being significantly faster than SotA models such as RAFT-stereo [18], GA-Net [45] and LEA-Stereo [3].