search

sovit-123 / fasterrcnn-pytorch-training-pipeline

PyTorch Faster R-CNN Object Detection on Custom Dataset
MIT License
211 stars 69 forks source link

Why are images without objects removed? #26

Open tolsicsse opened 1 year ago

tolsicsse commented 1 year ago

Thanks for this light weight faster rcnn implementation! However, I have been using Detectron2, and it is able to train on images without objects. I think it would be good to use them as negative examples.

sovit-123 commented 1 year ago

@tolsicsse Thanks for the suggestion. I think Darknet also retains the images without ground truth bounding boxes. I did not know that Detectron 2 also does that. I have been trying to figure out the pipeline for that for some time now. But generally, the albumentations augmentations part will throw an error without ground truth boxes. Many other libraries also discard images without objects. But I am sure that what you are suggesting is bound to help. Still I will try to implement it.