Train AdaptiveOcc with 4 RTX3090 GPUs
./tools/dist_train.sh ./projects/configs/adaptiveocc/adaptiveocc.py 4 ./work_dirs/adaptiveocc
Train AdaptiveOcc with 1 RTX3090 GPUs
python ./tools/train.py ./projects/configs/adaptiveocc/adaptiveocc.py --work-dir ./work_dirs/output --deterministic --no-validate
Eval AdaptiveOcc with 4 RTX3090 GPUs
./tools/dist_test.sh ./projects/configs/adaptiveocc/adaptiveocc_inference.py ./path/to/ckpts.pth 4
Eval AdaptiveOcc with 1 RTX3090 GPUs
python ./tools/test.py ./projects/configs/adaptiveocc/adaptiveocc_inference.py ./path/to/ckpts.pth --deterministic --eval bbox
Visualize occupancy predictions, occupancy groundtruth and the multi-scale occupancy groundtruth:
First, you need to generate prediction results. Here we use whole validation set as an example.
./tools/dist_test.sh ./projects/configs/adaptiveocc/adaptiveocc_inference_vis.py ./path/to/ckpts.pth 4
# python ./tools/test.py ./projects/configs/adaptiveocc/adaptiveocc_inference_vis.py ./path/to/ckpts.pth --deterministic --eval bbox
You will get prediction results in './visual_dir'. You can directly use meshlab to visualize .ply files or run visual_octree.py to visualize raw .npy files with mayavi:
python ./tools/visual_octree.py visual_dir/$npy_path$
Visualize multi-scale occupancy groundtruth:
python ./tools/visual_octree.py visual_dir/$npy_path$ --is_gt
Visualize occupancy groundtruth:
python ./tools/visual.py $npy_path$