/bstld

Sample scripts for the Bosch Small Traffic Lights Dataset

Primary LanguagePythonMIT LicenseMIT

Bosch Small Traffic Lights Dataset

This repository contains some scripts to get started with the Bosch Small Traffic Lights Dataset (BSTLD). Contributions are very welcome. Simply create a pull request.

Dataset

The dataset can be downloaded here. A preview of the dataset is available on YouTube by clicking on the image.

BSTLD Preview

Instructions on how to unzip *.zip.00X files can, for example, be found at https://hiro.bsd.uchicago.edu/node/3168 Update label files are in the label_files folder.

To convert Bosch Small Traffic Lights Dataset Annotations to Pascal VOC Format

 python bosch_to_pascal.py input_yaml out_folder

Sample Detections

A sample detection based on an adapted Yolo v1 model run on crops can be viewed at Sample Detector View]

Results

Method Execution time weighted mAP mAP Off Green Yellow Red External data Link
Baseline <100 ms 0.36 no https://ieeexplore.ieee.org/document/7989163/
Hierarchical Deep Architecture ~150 ms 0.53 no https://arxiv.org/abs/1806.07987
SSD Mobilenet V1 38 ms 0.60 0.41 0.00 0.68 0.41 0.55 no https://github.com/bosch-ros-pkg/bstld/blob/master/tf_object_detection/configs/ssd_mobilenet_v1.config
Faster RCNN NAS-A ~1560s 0.65 0.43 0.00 0.71 0.33 0.66 no https://github.com/bosch-ros-pkg/bstld/blob/master/tf_object_detection/configs/faster_rcnn_nas.config

Values are self-reported. The evaluation is performed on the test-set without empty frames. For different goals, e.g. using minimal training data, using external data only, or others, new tables can be created. We specifically encourage non-conventional approaches. Please make sure not to incorporate the test-set into your training, which includes multiple evaluations for different checkpoints of the same method. We understand that there can be larger variations between the different class average precisions, specifically due to the biased distribution. We will try to incorporate variations in results of the same method once reported.

Citation

In case of publication based on this dataset, please cite
@inproceedings{behrendt2017deep,
  title={A deep learning approach to traffic lights: Detection, tracking, and classification},
  author={Behrendt, Karsten and Novak, Libor and Botros, Rami},
  booktitle={Robotics and Automation (ICRA), 2017 IEEE International Conference on},
  pages={1370--1377},
  year={2017},
  organization={IEEE}
}