MS-DAYOLO

This is the official implementation of Multiscale Domain Adaptive YOLO (MS-DAYOLO) for Cross-Domain Object Detection based on Darknet (YOLOv4).

Installation

Please follow the instruction in AlexeyAB/darknet to install the requirements to compile Darkent and use YOLOv4 object detector.

Example Usage

Please follow these steps to adapt from Cityscapes to Foggy Cityscapes:

1. Run make to compile Darknet with MS-DAYOLO.
2. Download the pre-trained weights-file: yolov4.conv.137.
3. Download the required data file Cityscaples2Foggy.zip which has:

• source set with ground-truth annotation
• target set (with dummy annotation)
• test target set with ground-truth annotation.

5. Extract the file in data/ folder.
6. Train MS-DAYOLO:

./darknet detector train data/c2f.data cfg/ms-dayolo.cfg yolov4.conv.137 -dont_show -map -da

During training, the best weight will be saved in backup/ folder.

6. If the best weight file does not be updated for a long time, stop the training.
7. After training is finished, evaluate the trained MS-DAYOLO:

./darknet detector map data/c2f.data cfg/ms-dayolo.cfg backup/ms-dayolo_best.weights

8. For comparison with the original YOLOv4, train it:

./darknet detector train data/c2f.data cfg/yolov4.cfg yolov4.conv.137 -dont_show -map

and evaluate the trained YOLOv4:

./darknet detector map data/c2f.data cfg/yolov4.cfg backup/yolov4_best.weights

Results & Pretrained Weights

Adaptation from Cityscapes to Foggy Cityscapes

Method	mAP	Pretrained Weights
YOLOv4	35.64	download
MS-DAYOLO	43.04	download

Citation

@INPROCEEDINGS{ms-dayolo,
  author={Hnewa, Mazin and Radha, Hayder},
  booktitle={2021 IEEE International Conference on Image Processing (ICIP)}, 
  title={Multiscale Domain Adaptive Yolo For Cross-Domain Object Detection}, 
  year={2021},
  pages={3323-3327}}