[ECCV 2024] T-MAE: Temporal Masked Autoencoders for Point Cloud Representation Learning

---

🏠 About

This repository contains the official implementation of the paper T-MAE: Temporal Masked Autoencoders for Point Cloud Representation Learning by Weijie Wei, Fatemeh Karimi Najadasl, Theo Gevers and Martin R. Oswald.

🔥 News

• [2024/09/19] The code will be released soon.
• [2024/09/22] Release the code of evaluation on ONCE dataset.
• [2024/09/25] Training code on ONCE dataset released as well as the pretrained and finetuned weights.

Table of Contents

TODO

• Release ONCE evaluation code.
• Release ONCE training code.
• Release Waymo training code and inference code.

Installation

We test this environment with NVIDIA A100 GPUs and Linux RHEL 8.

conda create -n t-mae python=3.8
conda activate t-mae
conda install -y pytorch==1.11.0 torchvision==0.12.0 torchaudio=0.11.0 cudatoolkit=11.3 -c pytorch
conda install -y -c fvcore -c iopath -c conda-forge fvcore iopath
pip install "git+https://github.com/facebookresearch/[email protected]"
pip install numpy==1.19.5 protobuf==3.19.4 scikit-image==0.19.2 spconv-cu113 numba scipy pyyaml easydict fire tqdm shapely matplotlib opencv-python addict pyquaternion awscli open4d pandas future pybind11 tensorboardX tensorboard Cython 
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.11.0+cu113.html

pip install pycocotools
pip install SharedArray
pip install tensorflow-gpu==2.5.0
pip install protobuf==3.20

git clone https://github.com/codename1995/T-MAE
cd T-MAE && python setup.py develop --user
cd pcdet/ops/dcn && python setup.py develop --user

Data Preparation

Please follow the instruction of OpenPCDet to prepare the dataset. For the Waymo dataset, we use the evaluation toolkits to evaluate detection results, where the compute_detection_metrics_main file comes from Waymo-open-dataset API (Mar 2023) and its source code is C++ based.

data
│── waymo
│   │── ImageSets/
│   │── raw_data
│   │   │── segment-xxxxxxxx.tfrecord
│   │   │── ...
│   │── waymo_processed_data
│   │   │── segment-xxxxxxxx/
│   │   │── ...
│   │── waymo_processed_data_gt_database_train_sampled_1/
│   │── waymo_processed_data_waymo_dbinfos_train_sampled_1.pkl
│   │── waymo_processed_data_infos_test.pkl
│   │── waymo_processed_data_infos_train.pkl
│   │── waymo_processed_data_infos_val.pkl
│   │── compute_detection_metrics_main
│   │── gt.bin
│── once
│   │── ImageSets/
│   │── data
│   │   │── 000000/
│   │   │── ...
│   │── gt_database/
│   │── once_dbinfos_train.pkl
│   │── once_infos_raw_large.pkl
│   │── once_infos_raw_medium.pkl
│   │── once_infos_raw_small.pkl
│   │── once_infos_train.pkl
│   │── once_infos_val.pkl
│── ckpts
│   │── once_tmae_weights.pth
│   │── ...

Training & Testing

ONCE dataset

# t-mae pretrain & finetune on ONCE
bash scripts/once_train.sh

# Load provided pretrained model and finetune on ONCE
bash scripts/once_finetune_only.sh

# test
bash scripts/once_test.sh

Results

Waymo

Reproduced results to be updated soon. We could not provide the pretrained weights due to Waymo Dataset License Agreement.

ONCE

	mAP	Vehicle	Pedestrian	Cyclist	Weights
T-MAE (Pretrained)	-	-	-	-	once_tmae_pretrained.pth
T-MAE (Finetuned)	67.41	77.53	54.81	69.90	once_tmae_weights.pth

Citation

If you find this repository useful, please consider citing our paper.

@inproceedings{wei2024tmae,
  title={T-MAE: Temporal Masked Autoencoders for Point Cloud Representation Learning},
  author={Weijie Wei, Fatemeh Karimi Najadasl, Theo Gevers and Martin R. Oswald},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024}
}

Acknowledgements

This project is mainly based on the following repositories:

• GD-MAE
• SST
• OpenPCDet

We would like to thank the authors for their great work.