Awesome

Spatial-Temporal Enhanced Transformer Towards Multi-Frame 3D Object Detection

Implementation of paper: "Spatial-Temporal Enhanced Transformer Towards Multi-Frame 3D Object Detection". The implementation is mainly based on an Efficient, Flexible, and General deep learning framework, namely EFG.

Overview

• Installation
• Data
• Get Started
• Main Results
• Citation

1. Installation

1.1 Prerequisites

• gcc 5 - 7
• python >= 3.6
• cuda >= 10.2
• pytorch >= 1.6

# spconv
spconv_cu11{X} (set X according to your cuda version)

# waymo_open_dataset
## python 3.6
waymo-open-dataset-tf-2-1-0==1.2.0

## python 3.7, 3.8
waymo-open-dataset-tf-2-4-0==1.3.1

1.2 Build from source

git clone https://github.com/Eaphan/STEMD.git
cd STEMD
pip install -v -e .
# set logging path to save model checkpoints, training logs, etc.
echo "export EFG_CACHE_DIR=/path/to/your/logs/dir" >> ~/.bashrc

2. Data

2.1 Data Preparation - Waymo

# download waymo dataset v1.2.0 (or v1.3.2, etc)
gsutil -m cp -r \
  "gs://waymo_open_dataset_v_1_2_0_individual_files/testing" \
  "gs://waymo_open_dataset_v_1_2_0_individual_files/training" \
  "gs://waymo_open_dataset_v_1_2_0_individual_files/validation" \
  .

# extract frames from tfrecord to pkl
CUDA_VISIBLE_DEVICES=-1 python cli/data_preparation/waymo/waymo_converter.py --record_path "/path/to/waymo/training/*.tfrecord" --root_path "/path/to/waymo/train/"
CUDA_VISIBLE_DEVICES=-1 python cli/data_preparation/waymo/waymo_converter.py --record_path "/path/to/waymo/validation/*.tfrecord" --root_path "/path/to/waymo/val/"

# create softlink to datasets
cd /path/to/STEMD/datasets; ln -s /path/to/waymo/dataset/root waymo; cd ..
# create data summary and gt database from extracted frames
python cli/data_preparation/waymo/create_data.py --root-path datasets/waymo --split train --nsweeps 4
python cli/data_preparation/waymo/create_data.py --root-path datasets/waymo --split val --nsweeps 4

2.2 Data Preparation - nuScenes

# create softlink to datasets
cd /path/to/STEMD/datasets; ln -s /path/to/nuscenes/dataset/root nuscenes; cd ..
python cli/data_preparation/nuscenes/create_data.py --root-path datasets/nuscenes --version v1.0-trainval --nsweeps 31  # 1 sample frame + 30 sweeps frame (1.5s)

3. Get Started

3.1 Training & Evaluation

cd playground/detection.3d/waymo/stemd/STEMD.waymo.resnet18.cdn.epoch12

efg_run --num-gpus x  # default 1
efg_run --num-gpus x task [train | val | test]
efg_run --num-gpus x --resume
efg_run --num-gpus x dataloader.num_workers 0  # dynamically change options in config.yaml

Models will be evaluated automatically at the end of training. Or,

efg_run --num-gpus x task val

4. Main Results

All models are trained and evaluated on 8 x NVIDIA A100 GPUs.

Waymo Open Dataset - 3D Object Detection (val L2- mAP/mAPH)

5. Citation

@misc{zhu2023efg,
    title={EFG: An Efficient, Flexible, and General deep learning framework that retains minimal},
    author={EFG Contributors},
    howpublished = {\url{https://github.com/poodarchu/efg}},
    year={2023}
}
@article{zhang2023spatial,
  title={Spatial-Temporal Enhanced Transformer Towards Multi-Frame 3D Object Detection},
  author={Zhang, Yifan and Zhu, Zhiyu and Hou, Junhui and Wu, Dapeng},
  journal={arXiv preprint arXiv:2307.00347},
  year={2023}
}
@inproceedings{zhu2023conquer,
  title={Conquer: Query contrast voxel-detr for 3d object detection},
  author={Zhu, Benjin and Wang, Zhe and Shi, Shaoshuai and Xu, Hang and Hong, Lanqing and Li, Hongsheng},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={9296--9305},
  year={2023}
}