Image processing with Halide
summary
Zivid rewrites an OpenCL structured‑light pipeline in Halide and keeps the same algorithm while moving scheduling into the library, cutting a baseline CPU step from 61 ms → 0.6 ms (≈100×) by adding compute_at, vectorize, parallel and GPU tiling; code lives in Python bindings (C++ available) and compiles ahead‑of‑time to object files, supports Halide in C++/Python with AOT and extern calls for algorithms that don’t fit the pure feed‑forward model, and notes one pain point: missing Python static typing compared with C++.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:29Introduction
- 5:18Credit where credit is due
- 7:26Humanlike vision
- 11:09Patterns
- 13:10What is Halide
- 21:34Driving example
- 25:09Image params
- 29:44For Minimum
- 31:04Projector Column
- 34:59Visualization
- 36:09Compute at root
- 39:59Vectorization
- 45:14Auto schedulers
- 46:02Rewriting the pipeline
- 50:34Conclusion
- 52:23Questions
related talks
49mGraph Based Processing of Big Images
47mHow we made Zivid Two - our 2nd generation 3D machine vision camera for robotics
01h 27m“Patterns and Techniques Used in the Houdini 3D Graphics Application ”
29mReconstructing Unseen Dimensions in Biophotonics Through Python
45mQt, multimedia, and computer vision – a great match for Embedded Linux
43mOh, Oh, Oh, It's Magic!
31mPainting With Light
15mFast Lossless Depth Image Compression