bayesian-yolov3

YOLOv3 object detection architecture with uncertainty estimation in TensorFlow.

Accompanying code for: https://arxiv.org/abs/1905.10296 (IEEE Xplore: https://ieeexplore.ieee.org/document/8917494/)

Citation

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

@inproceedings{kraus_uncertainty_2019,
  address = {Auckland, New Zealand},
  title = {Uncertainty {Estimation} in {One}-{Stage} {Object} {Detection}},
  url = {https://ieeexplore.ieee.org/document/8917494/},
  doi = {10.1109/ITSC.2019.8917494},
  booktitle = {2019 {IEEE} {Intelligent} {Transportation} {Systems} {Conference} ({ITSC})},
  publisher = {IEEE},
  author = {Kraus, Florian and Dietmayer, Klaus},
  month = oct,
  year = {2019},
  pages = {53--60}
}

Notes

• Training examples with documentation:

• pretraining.py - pretraining for models with uncertainty estimation

• uncertainty_training.py - training for models with uncertainty estimation, use checkpoints produced by pretraining.py as a starting point.

• yolov3_training.py - standard yolov3 without any uncertainty estimation

• look for "edit" comments

• Forward passes:

• detect.py (processes a list of images)

• inference_*.py scripts. They process tfrecord files and produce ECP (euro city persons) formated json files.

• Note NMS yields up to 1000 boxes, might be slow. Change the "nms" functions if you want better performance.

• Current NMS implementation ignores classes. Example code used in the paper is given as comments (only works for two classes).

• look for "edit" comments

• tfrecords format:

• same as for the tensorflow object detection API, however we also support tfrecords files where the label ids start at 0 instead of 1. This is controlled by setting "implicit_background_class" to True (start at 1) or False (start at 0).

• example script to create tfrecordsfile is provided (create_tf_records_citypersons.py)

• Pretrained yolov3 weights:

• you need to download the "darknet53.conv.74" from the original yolov3 site (pjreddie).

• Most things you can change should be marked with an "edit" comment.