Implicit Tracking

This repository is the official implementation of the paper:

Online Adaptation for Implicit Object Tracking and Shape Reconstruction in the Wild
Jianglong Ye, Yuntao Chen, Naiyan Wang, Xiaolong Wang
RA-L 2022, IROS 2022

Project Page / ArXiv / Video

Environment Setup

(Our code has been tested with python 3.8, torch 1.8.0, CUDA 11.1 and RTX 3090)

To set up the environment, follow these steps:

conda create -n itrack python=3.8 -y && conda activate itrack
conda install pytorch==1.8.0 torchvision==0.9.0 torchaudio==0.8.0 cudatoolkit=11.1 -c pytorch -c conda-forge -y

pip install "git+https://github.com/facebookresearch/pytorch3d.git@v0.5.0"
pip install waymo-open-dataset-tf-2-3-0 protobuf==3.20.0  # for data processing, other version may not work
pip install matplotlib yacs opencv-python scikit-image scikit-learn trimesh numba shapely pandas webdataset numpy==1.23.5 gdown

We also provide a environment.yml file for reference.

Data Preparation

Data Structure

The data should be organized as follows:

PROJECT_ROOT
└── data
    |-- waymo
    |   |-- raw
    |   |   └── validation
    |   |       |-- segment-10203656353524179475_7625_000_7645_000_with_camera_labels.tfrecord
    |   |       ...
    |   |-- processed
    |   |   └── validation
    |   |       |-- segment-10203656353524179475_7625_000_7645_000_with_camera_labels.tar
    |   |       ...
    |   └── splits
    |       |-- easy_list.json
    |       ...
    └── kitti
        |-- raw
        |   |-- data_tracking_velodyne.zip
        |   |-- data_tracking_label_2.zip
        |   ...
        └── processed
            └── training
                |-- clean_pcs
                |   |-- 0019
                |   ...
                |-- label_02
                ...

Waymo Open Dataset

Download the perception dataset (v1.2) from the official website and organize it as described above.

It's recommended to use gcloud to download the data, here are example steps to install gcloud on Ubuntu 20.04 and download the data (please sign the license agreement on the website first):

curl -O https://dl.google.com/dl/cloudsdk/channels/rapid/downloads/google-cloud-cli-412.0.0-linux-x86_64.tar.gz
tar -xf google-cloud-cli-412.0.0-linux-x86_64.tar.gz
./google-cloud-sdk/install.sh  # follow the instructions to install

# open a new terminal so that the changes take effect
gcloud init  # log in with the google account

# download the data, around 191GB
mkdir -p data/waymo/raw
gcloud storage cp -r gs://waymo_open_dataset_v_1_2_0_individual_files/validation data/waymo/raw/

After download, run the following command to process the data:

# cd PROJECT_ROOT
python tools/prepare_waymo.py

KITTI Tracking Dataset

Download data from the official website and our detection results from here. Organize the data as described above.

Here are example steps to download the data (please sign the license agreement on the website first):

mkdir data/kitti/raw && cd data/kitti/raw
wget -c https://s3.eu-central-1.amazonaws.com/avg-kitti/data_tracking_velodyne.zip  # around 34 GB
wget -c https://s3.eu-central-1.amazonaws.com/avg-kitti/data_tracking_label_2.zip
wget -c https://s3.eu-central-1.amazonaws.com/avg-kitti/data_tracking_calib.zip

# detection results for scenes 19, 20
gdown --fuzzy https://drive.google.com/file/d/12EXXKtv8FMDJ_z0YnaaBNiRJzF3iKJft/view?usp=share_link
# gdown may not work due to the limit of Google Drive, please download it manually if necessary

After download, run the following command to process the data:

# cd PROJECT_ROOT
python tools/prepare_kitti.py

SOT

We provide shape models pre-trained on the ShapeNet dataset here, please download and put them under ckpts. Here are example commands to use gdown to download the models (gdown may not work due to the limit of Google Drive, please download them manually if necessary):

cd ckpts
gdown --fuzzy https://drive.google.com/file/d/1UKRVgcoNhbdCn0xsBA96YD0ehYS3o0ny/view?usp=share_link
unzip waymo.zip && rm waymo.zip

gdown --fuzzy https://drive.google.com/file/d/18N7UPlu-CAYT8XSt_G8Vs5iBCXSSbgVH/view?usp=share_link
unzip kitti.zip && rm kitti.zip

Run following command to perform SOT:

# cd PROJECT_ROOT
export PYTHONPATH=.
python tools/sot.py --config-file ./configs/waymo.yaml
python tools/sot.py --config-file ./configs/kitti_det.yaml

Calculate the metrics:

python tools/evaluate.py --exp-dir ./output/waymo/summary
python tools/evaluate.py --exp-dir ./output/kitti_det/summary

The results should be similar to:

# waymo
success: 62.70
precision: 66.13

# kitti_det
success: 62.14
precision: 77.65

Note that the results may be slightly different due to the randomness in the optimization process.

Citation

If you find this work useful in your research, please consider citing:

@article{ye2022online,
  author    = {Ye, Jianglong and Chen, Yuntao and Wang, Naiyan and Wang, Xiaolong},
  title     = {Online Adaptation for Implicit Object Tracking and Shape Reconstruction in the Wild},
  journal   = {IEEE Robotics and Automation Letters},
  year      = {2022},
}

jianglongye / implicit-tracking