Reproduce “Anytime Stereo Image Depth Estimation on Mobile Devices” on OpenResearch

About this paper

Yan Wang, Zihang Lai, Gao Huang, Brian H. Wang, Laurens van der Maaten, Mark Campbell, Kilian Q. Weinberger

Researchers at Cornell University, University of Oxford, and Facebook AI Research developed AnyNet, a multi-stage, end-to-end learned architecture for stereo depth estimation that achieves real-time performance on mobile devices. The method provides dynamic trade-offs between computation time and accuracy, enabling up to 35 FPS on an NVIDIA Jetson TX2 while maintaining competitive accuracy with significantly fewer parameters than state-of-the-art models.

Key insights

• Anytime prediction can be effectively applied to structured prediction tasks like disparity estimation by employing a progressive refinement approach using residual predictions across multiple resolution stages.
• Computing residual disparity maps at higher resolutions, rather than full maps, drastically reduces computational cost by limiting the search range in the cost volume.
• A lean architecture with an exceptionally low parameter count (40,000) is achievable without sacrificing competitive accuracy, making models viable for deployment on severely constrained hardware.

Summary by alphaXiv.