SeeAnythingFar

Teaching Robots To See Far Away Obstacles!

The idea behind doing this was long range and sparse detection of 3D pointclouds to do long horizon planning for autonomous vehicles, which have heavy mass or are moving at high speed, as they cannot stop immediately, for which long horizon planning has to be done for which we need to detect, track and segment faraway objects.

Architecture

Slurm

• added local cuda setup for version 12.8 for compatibility
• code is in slurm/neon/cuda_setup.sh

OpenPCDet

• currently I have used it as a backbone
• try to install openmmlab environment from their official website

# some additional dependency
conda activate openmmlab
pip install spconv-cu120

# for visualizations
pip install open3d

• setup the repo using

cd /src/seeanythingfar/openPCDet/
python3 setup.py develop

Pretained weights

• currently used pv_rcnn weights for 3D object detection for far distance from link
• a copy of the orginal semantic-kitti dataset is there on the neon server

python3 demo.py \ 
--cfg_file cfgs/kitti_models/pv_rcnn.yaml \ 
--ckpt pv_rcnn_8369.pth \
--data_path /scratch/soumo_roy/semantic-kitty-dataset/dataset/sequences/00/velodyne/000000.bin 

Slide Deck

• I am attaching the slide deck of the experiments done link

Environment setup

uv pip install -e .
source .venv/bin/activate