Reproducing results on segmentation using training from scratch

[sonamsingh19]

Hi,
I tried to dino vit small from scratch for some research on datasets like mapillary and imagenet (inspired from dinov2 dataset picking) but in linear probing, it learns large classes like vehicles but ignores traffic lights with iou 0. Any insights on why this might be the case?