Hi all,
I work with RGB-D cameras and kept hitting the same thing. On glass and clear plastic the camera returns confident, wrong depth, with no error raised.
On TransCG data, 68.6% of the glass pixels an L515 returned were wrong by more than 5% of true depth, and nearly all read too far, so the space occupied by glass looks free.
limit_case_ros2 is a composable node that detects suspect basins in the depth image (Priority-Flood on inverted depth) and publishes filtered depth plus a mask of flagged pixels. It offers:
- 7.4 ms per frame at 1280x720 (core filter)
- a demo rosbag and eval scripts that reproduce every number in the README
- an experimental suspected_obstacles cloud (the Nav2 demo came out negative, and the README says so)

One thing I cared about: the repo reports costs and failures next to the wins, including a pre-registered held-out test of my own v0.2 method that failed and is published in full.
I’ve tested it on Jazzy and Humble (Apache-2.0).
I’d welcome feedback, especially from anyone who has fought glass or clear plastic on a real robot.
If you have a rosbag with transparent objects in frame, I’ll run the filter on it and send back what it flags.