RobDet3D

This repository is deeply inspired by OpenPCDet and MMDetection3D, and incorporates their advantages. In this repo, we develop HAVSampler and GridBallQuery to speed up poin-based model and its deployment.

INSTALL

pip install -r requirements.txt
pip install spconv-cuxxx(e.g.cu113)
python setup.py develop

pre-train model can be downloaded from Google Drive.

Dataset

refer to OpenPCDet to prepare your data:

mkdir data && cd data
ln -s path/to/your/kitti/dataset kitti
cd ../
python -m rd3d.datasets.kitti.kitti_dataset \
create_kitti_infos configs/base/datasets/kitti_3cls.py

Training

train your model like this:

python tools/train.py \
--cfg configs/iassd/iassd_hvcsx1_4x8_80e_sparse_kitti_3cls.py \
wandb --group rd3d

or train with multi-GPU

accelerate launch tools/train.py \
--cfg configs/iassd/iassd_hvcsx1_4x8_80e_sparse_kitti_3cls.py \
wandb --group rd3d

Export ONNX Model

python tools/deploy/export_onnx.py                                      \
--cfg_file configs/iassd/iassd_hvcsx1_4x8_80e_kitti_3cls\(export\).py   \
--ckpt tools/models/iassd_hvcsx1_4x8_80e_kitti_3cls\(export\).pth       \
--onnx tools/models/trt

Then the exported onnx model can be found in fold tools/models/trt.

Export TensorRT Model

python tools/deploy/export_trt.py                                         \
--onnx ./tools/models/trt/iassd_hvcsx1_4x8_80e_kitti_3cls\(export\).onnx  \
--batch 1                                                                 \
--type FP32  # FP32 / FP16 / INT8

Test

python tools/deploy/test_trt_viz.py \
--engine ./tools/models/trt/iassd_hvcsx1_4x8_80e_kitti_3cls\(export\).engine

if susses, the following figure can be seen in the visualization window.

Profile

python tools/deploy/trt_profile.py                                            \
--engine tools/models/trt/iassd_hvcsx1_4x8_80e_kitti_3cls\(export\).engine    \
--batch 1                                                                     \
--build_in  # or not

And the following results will be shown in terminal:

name	layers	average (ms)	median (ms)	percentage (%)
HAVSampling	1	0.106	0.102	0.9
BallQuery	8	5.197	5.054	42.4
ForeignNode	8	4.144	4.079	34.2
NMSBEV	1	0.137	0.137	1.1
Others	48	2.402	2.358	19.8
Total	66	12.0	11.9	100.0

OuyangJunyuan/RobDet3D