The AIEA lab @ UCSC for the Lidar Scene Flow 2026 challenge
Documentation on function available in docdumentation.md
sh setup.sh 0
and terminate with
sh setup.sh 1
The datasets are stored in an 8TB pvc called 'lidarflow' in the aiea-slugbotics namespace.
In the PVC run:
. start.sh
to set up the necessary environments for development/training. It is about as cuda version agnostic as possible, and should train and install the correct libraries. Training by default is with FP16. BF16 might produce some better results.
Jobs are available by running
sh setup.sh 3
and terminated with
setup.sh 4
The datasets can be downloaded with the shell script in the
/downloadDataset directory
git cloning and running:
bash resumeDownload.sh
will download the dataset and resume if there is an unexpected connection loss. The screen command also should work but for some reason causes pods to frequently fail with exit code 137, so it might be causing an out of memory error on pods that are low on memory anda have high output to stdout.