@ML-GOD

Many thanks for sharing this. I haven't delved too deeply into the dataset, but what I've noticed so far is that the error appears to be a 'bug' in the code for generating the polygons, not necessarily intentional. The shift is consistent for most of the images, especially the drone ones. The values you've shared are rescaled to 512, correct? I also have a value for the shift, however, a few samples have slightly different shift values. Also, not all images/polygons have this problem. The challenge is how to identify which of the images/polygons to apply corrections to. So far, my observation is about 1 in (10-20) images have this error. Just like you, I wish I had the time to work on this actively.

Ya. I have mentioned the ids of wrong polygons in the notebook. I tried to work on segmentation, but data was so bad that I had to work on data first. It took me only 3 hrs for this. Also, train the model around 40-50 epochs to get the best results

Thanks. I will try it out. How did you do it in 3hrs? Physical inspection per image or what?

Its done. You can now view the notebook