Show HN: Rebuilding ARKit 3D Reconstruction

Category: ai-ml

Tags: 3d-reconstruction, robotics, slam, dataset, data-pipeline

Score: 6.3/10 (Innovation: 6, Technical: 7, Documentation: 5, Utility: 6)

This project is an open-source attempt to rebuild Apple's ARKit 3D reconstruction pipeline for generating robot training data, using the ARKitScenes dataset as ground truth. It focuses on replacing closed-source components like visual-inertial SLAM, depth prediction, and depth upsampling with open models such as SimpleRecon and Prompt Depth Anything, storing data in Rerun's open data layer. The detailed write-up explains the challenges of heterogeneous sensor streams and the importance of a unified data format for iterative robot learning.

Target audience: robotics researchers and engineers, computer vision developers, data engineers

Repository: https://twitter.com/pablovelagomez1/status/2082256328966422542

View on Hacker News