How does your dataset_create.py works?

[tussank]

Hey, I am trying to train a fashion Object Detector. According to the steps given in your readme file, I first have to use dataset_create.py Now as per my understanding that creates many cropped images out of each image via selective search. So, after running that script, in my dataset folder, in train, test and val, there are many cropped images. Many of these images have just the head of the person wearing those clothes and many other wrong images. So, won't such images confuse our object detector when we train? I need your help to understand how this selective search works and how it is useful?

Thanks

[hassanmasood1]

The author here is creating dataset where the IOU is less than 1.0. If you notice the name of the image holds IOU value, so it will help model learn examples where the IOU values are low BUT i think for cropped images where IOU are almost 0 we are still assigning label which will confuse the model. There should be a check somewhere to drop label for cropped images where IOU value is less than some threshold (something like <0.7 IOU)