search

sovit-123 / fasterrcnn-pytorch-training-pipeline

PyTorch Faster R-CNN Object Detection on Custom Dataset
MIT License
223 stars 75 forks source link

Modify mosaic #34

Closed sovit-123 closed 1 year ago

sovit-123 commented 1 year ago
  1. Non-mosaic training now carry out aspect ratio resizing. Better performance than square training. Can use --square-training for square resizing.
  2. Mosaic augmentation according to Ultralytics YOLOv5 now. But requires a lot more data. Smaller models tend to perform less effectively now. Can do square-training along with mosaic which helps smaller models also.
  3. With aspect ratio resizing and mosaic (default training settings) larger models can be trained much longer without overfitting.
  4. A few other minor changes.