NVIDIA/ncore v19.8.4

on GitHub

Highlights

Bug fix release.

• OpenCVFisheyeCameraModel.image_points_to_camera_rays now returns the exact inverse of the fisheye projection for all calibrations. It previously ran three unguarded Newton iterations from an equidistant starting point. For strongly distorted calibrations the result could be off by more than 1 rad, land on the decreasing branch of the polynomial, or be NaN. The inverse now falls back to bracketed Newton iterations where plain ones don't converge. The newton_iterations default changed from 3 to 20 and is now an upper bound, so well-conditioned calibrations cost the same as before. Image points beyond the maximum of the forward polynomial have no inverse and now map to the boundary ray of its increasing branch.

Changes since v19.8.3

🪲 Fixed

• (sensors) bracket the OpenCV fisheye inverse to the monotonic branch (8b7eac6) - @janickm

⚙️ CI

• publish the documentation built by build-and-test (26cb447) - @janickm
• refresh the Bazel disk cache on every push (0128cbe) - @janickm
• remove merge queue triggers (d17de7e) - @janickm
• release notes tool in Python, commit check in build-and-test (831e4bc) - @janickm
• separate the changes since the previous pre-release in release notes (ce4de7c) - @janickm

🏗️ Build

• don't compress the repacked patched torch wheel (73d9376) - @janickm
• pinned cocogitto via rules_multitool (056c0b7) - @janickm

🧪 Tests

• (sensors) share the lidar model between tests of a parameterization (194f2f7) - @janickm
• run tests with a single OpenMP thread (913076b) - @janickm

Full diff: v19.8.3...v19.8.4