3DSSD-torch

This is the 3DSSD implementation rewritten based off OpenPCDet framework. The code is orginized by the guideline given in the official repo as the figure below:

• Update 5/12/2021: added 3D IoU loss
• Update 4/11/2021: added pointpainting
• Update 3/09/2021: supported multi-class prediction

Requirements

The codes are tsted under the following environment:

• Ubuntu 20.04
• Python 3.7
• Pytorch 1.7.1
• CUDA 11.1

Installation

The procedure is identical to the office installation guide of OpenPCDet, you can also refer to the following:

1. Install required dependency by running pip install -r requirements.txt.

2. Refer to the link to install spconv (although not used in the repo, install for the ease of not modifying the original codebase)

3. Install rotated IoU operator. Go to pcdet/ops/Rotated_IoU/cuda_ops folder, run the command python setup.py develop.

4. Install the OpenPCDet related libraries by running python setup.py develop

5. Preprocess KITTI dataset by running the following command from project directory: python -m pcdet.datasets.kitti.kitti_dataset create_kitti_infos tools/cfgs/dataset_configs/kitti_dataset.yaml

6. (Optional) If training with pointpainting, painted dataset should first be generated by python -m pcdet.datasets.kitti.paint_kitti from project root directory. After the dataset is generated, re-preprocess the ground truth database by python -m pcdet.datasets.kitti.kitti_dataset_painted create_kitti_infos

Visualization

Refer to visualization for visualization info.

Train

Train the 3DSSD model by running the command in tools directory:

python train.py --cfg_file cfgs/kitti_models/3dssd.yaml

Pretrained Model

The pretrained models are provided in the output folder under ckpt directory, you can examine the pretrained model from tools directory by:

python test.py --cfg_file cfgs/kitti_models/3dssd.yaml --ckpt ../output/kitti_models/3dssd/default/ckpt/checkpoint_epoch_120.pth

Performance

Models with IoU loss are considerably better than original model, Car, Pedestrain and Cyclist as follow all under R11 criteria, if you are interested in R40 result as well, refer to the output folder.

Car [email protected], 0.70, 0.70:
bbox AP:96.6985, 90.1025, 89.4729
bev  AP:90.0036, 88.3474, 86.8328
3d   AP:89.0494, 79.1208, 78.1524
aos  AP:96.66, 89.97, 89.25
Pedestrian [email protected], 0.50, 0.50:
bbox AP:71.7494, 67.7955, 63.0582
bev  AP:63.7641, 57.1950, 54.4950
3d   AP:58.7558, 53.5601, 49.4172
aos  AP:67.95, 63.76, 59.10
Cyclist [email protected], 0.50, 0.50:
bbox AP:94.4467, 77.8638, 75.6069
bev  AP:92.8921, 74.7190, 71.6883
3d   AP:90.9060, 72.0020, 66.5100
aos  AP:94.36, 77.25, 75.00

As a comparison, here is the result for original model with L1 box loss:

Car [email protected], 0.70, 0.70:
bbox AP:96.4641, 90.0299, 89.4182
bev  AP:89.9948, 87.9284, 85.8481
3d   AP:88.5553, 78.4563, 77.3031
aos  AP:96.43, 89.94, 89.25
Pedestrian [email protected], 0.50, 0.50:
bbox AP:72.2514, 69.3984, 64.1170
bev  AP:63.2845, 58.1731, 55.0167
3d   AP:58.1822, 54.3187, 49.5647
aos  AP:65.99, 62.96, 57.97
Cyclist [email protected], 0.50, 0.50:
bbox AP:94.4852, 82.3373, 77.1431
bev  AP:92.4290, 73.7010, 70.8207
3d   AP:86.2519, 70.4854, 65.3238
aos  AP:94.41, 81.76, 76.66

Acknowledgement

Many thanks to qiqihaer and his excellent work on reimplmentation of 3DSSD. I borrowed some code from his repo including part of the head and coder.

Also I refered to the code from MMdetection3D to make the code structure more organized.