sk-ram / YOLO-Fine

A Keras (TF Backend) based implementation of YOLO-Fine, a novel one-stage object detector for detecting vehicles from aerial data.

YOLO-Fine

Keras (TF Backend) implementation of YOLO-Fine

Read the paper: https://doi.org/10.3390/rs12152501

What is YOLO-Fine?

YOLO-Fine[1] is a novel one-stage detector for detecting and classifying small vehicles from aerial and satellite imagery.

Our dataset

We trained YOLO-Fine[1] on the VEDAI (https://downloads.greyc.fr/vedai/) dataset. Anchors were generated using k-means clustering, and can be found in the ./vedai_anchors.txt file. Classes were set to 0 to focus on bounding box classification.

Usage:

Use with sample weights:

Run ./yolo_image.py

python ./yolo_image.py

Train on your dataset:

Modify train.py with your:

• Annotations file
• Anchors file
• Classes file
• Input shape
• Training epochs
• Target batch size

Then run ./train.py

Note that the model will train with the darknet body frozen, and will then unfreeze those layers and continue training.

python ./yolo_image.py

Acknowledgments

We would like to thank the authors of the originial YOLO-Fine paper as well as @pjreddie and @qqwweee.

References

[1] Pham, Minh-Tan et al. “YOLO-Fine: One-Stage Detector of Small Objects Under Various Backgrounds in Remote Sensing Images.” Remote Sensing 12.15 (2020): 2501. Available: http://dx.doi.org/10.3390/rs12152501.