search

ultralytics / xview-yolov3

xView 2018 Object Detection Challenge: YOLOv3 Training and Inference.
https://docs.ultralytics.com
GNU Affero General Public License v3.0
256 stars 57 forks source link

Fine tune model for using on grayscale images of xview #13

Closed sawhney-medha closed 5 years ago

sawhney-medha commented 5 years ago

I want to fine tune this model on a different dataset that is grayscale. I understand that train.py has a resume option which will start training from previous weights. But train.py also uses target_c60.mat. I cannot understand how to create the .mat files for my own custom data. Please help

glenn-jocher commented 5 years ago

@sawhney-medha Hello, thank you for your interest in our work! Please note that this repository is no longer actively maintained. Instead we recommend all users start from github.com/ultralytics/yolov3 and adapt it to their specific needs.

The target_c60.mat file you mentioned can be read by scipy.io, however it was generated by a seperate MATLAB script, which mainly trims poor labels and creates a k-means anchors from the remaining good candidates. https://github.com/ultralytics/xview-yolov3/blob/9d30dabaeccf7937d2e66dca7324890d65e1799b/utils/analysis.m#L79

In order to adapt github.com/ultralytics/yolov3 to geospatial data similar to xview you mainly need to implement the sliding window inference and chip training, and ideally also do a k-means search of your anchors.

sawhney-medha commented 5 years ago

Okay, Thank you.

glenn-jocher commented 5 years ago

Great, you're welcome!