Repository navigation
Original RoboSense M1: pretrained detector transfer and minimum useful adaptation data #1776
Unanswered
TwelveInchPianist
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I'm evaluating OpenPCDet as a possible replacement for parts of a custom geometric lidar detector. I have not yet run an OpenPCDet model on these excerpts, so this is a compatibility/adaptation question rather than a bug report.
The sensor is an original RoboSense M1 (not Plus), 10 Hz, strongest single return, nonuniform forward MEMS scan pattern, nominal average 0.2-degree spacing. It is fixed in a recessed vehicle-roof position. I have native camera footage, vehicle motion/IMU and raw lidar from ordinary drives, but limited manually labeled 3D objects and no staged deer/pedestrian test course.
Four small review cases, raw PCAP excerpts, decoded XYZ/intensity/per-point timestamps, original camera segments and uncertainty notes: https://github.com/TwelveInchPianist/odyssey-lidar-samples
The objective is audible notice of pedestrians, cyclists, animals and credible unknown roadside/road objects, especially at night. Presence is useful even when class or collision path is unknown. The M1 can establish a cyclist track around 95 m in one daylight case; that does not establish recognition at that range. Ground/curb ambiguity, sparse targets and identity fragmentation remain problems.
Current inference target is a Z1 Extreme tablet; offline machines include Apple silicon and a Ryzen desktop without a strong NVIDIA GPU. I recognize CUDA/custom-op requirements may rule out direct deployment; that limitation is part of the decision, not something I want to hide. I would appreciate candid guidance or a pointer to an existing discussion.
All reactions