TSTTC for Time-To-Contact estimation.

Introduction

The TSTTC dataset is a large scale dataset for TTC estimation via monocular images. For more details, please refer to our paper on Arxiv. To download the dataset and evaluate result on test set, please refer to the TSTTC in baidu cloud.

Dataset Structure

The TSTTC dataset is organized into four main folders: train, val, test, and nerf, following the data structure outlined below:

TSTTC Dataset
├── train
│ ├── {bag_stamp1}.pkl
│ ├── {bag_stamp1}
│ │ ├── cam1
│ │ │ ├── 16xxxx.jpg
│ │ │ ├── 16xxxx.jpg
│ │ │ └── ...
│ │ ├── cam3
│ │ │ ├── 16xxxx.jpg
│ │ │ ├── 16xxxx.jpg
│ │ │ └── ...
│ │ └── ...
│ └── ...
├── val
│ └── ...
├── test
│ └── ...
└── nerf
  └── ...

Within each set, there are multiple bags identified by a unique bag stamp. Each bag consists of up to five camera subfolders: cam1, cam3, cam4, cam8, and cam9, except for the nerf part, which only contains cam1. The cam1, cam3, and cam4 subfolders correspond to frontal view cameras with different focal lengths, while cam8 and cam9 represent backward view cameras. The images within each camera subfolder are named based on the timestamp when they were captured. Additionally, each bag is associated with an annotation file named bag_stamp.pkl.

Annotation Format

The annotation files in the TSTTC dataset are formatted as a list of dictionaries. Each dictionary within the file represents the annotations for a specific camera, with the timestamp serving as the key and the frame annotations as the corresponding value. The frame annotation for each timestamp is represented as a list of objects. Each object is a dictionary containing the following properties:

Annotation File (List of Dictionaries)
├── Dict 1
│ ├── Timestamp 1
│ │ ├── Object 1
│ │ │ ├── bag_stamp
│ │ │ ├── box2d
│ │ │ ├── cam_id
│ │ │ ├── id
│ │ │ ├── occ_ratio
│ │ │ ├── same_lane
│ │ │ ├── ts
│ │ │ └── ttc_imu (Not available in test set)
│ │ ├── Object 2
│ │ │ ├── ...
│ │ ├── ...
│ ├── Timestamp 2
│ │ ├── ...
│ ├── ...
├── Dict 2
│ ├── ...
├── ...

Here is a detailed description of each property:

Property	Description
bag_stamp	A string representing the bag stamp of the object, indicating its association with a particular bag within the dataset.
box2d	A list of floats representing the 2D bounding box of the object in the image. In (x1,y1,x2,y2) format and coordinates are normalized to 0~1
cam_id	An identifier for the camera to which the object belongs.
id	An integer representing the tracking ID of the object across all cameras.
occ_ratio	A float value denoting the occlusion ratio of the object in the 2D images.
same_lane	A boolean flag indicating whether the object is in the same lane as the ego vehicle.
ts	An integer representing the timestamp of the object.
ttc_imu	A float value representing the ground truth TTC of the object. Not available in the test set.

Please note that the property ttc_imu is not available in the test set, as mentioned in the dataset description.

Dataset API

For the usage of the dataset API, please refer to TSTTCDatasetDemo.

Quick Start for Baseline Methods

Installation

Install TSTTC from source.

git clone https://github.com/tusen-ai/TSTTC
cd TSTTC

Create conda env.

conda create -n TSTTC python=3.7

conda activate TSTTC

pip install -r requirements.txt

cd cuda_ops

pip3 install -v -e .

cd ..

Evaluation

Evaluate the Pixel MSE.

python tools/eval_baseline.py --path [path_to_your_val_set]

Evaluate the Deep Scale

Step1. Replace the valset_dir and valAnnoPath of the exp_file to the path of your own validation set

Step2. Run the evaluation code

python tools/eval.py -f ./exp/Deep_TTC.py -c [path_to_your_weights] --path [path_to_your_val_set] -d 1 -b 8 --fp16 --box_level

Training

Step1. Replace the trainset_dir and trainAnnoPath of the [exp_file](./exp/Deep_TTC.py) to the path of your own training set

Step2. Run the training code

python tools/train.py -f ./exp/Deep_TTC.py -d 1 -b 8 --fp16 

Acknowledgements

Expand

• https://github.com/Megvii-BaseDetection/YOLOX

Reference

If you use TSTTC in your research, please cite our work:

@article{shi2023tsttc,
      title={TSTTC: A Large-Scale Dataset for Time-to-Contact Estimation in Driving Scenarios}, 
      author={Yuheng Shi and Zehao Huang and Yan Yan and Naiyan Wang and Xiaojie Guo},
      year={2023},
      journal={2309.01539},
}