ajtao
commented
3 years ago The centroid file is something that's created once for a given dataset for the purpose of uniform class sampling. Uniform sampling (enabled with class_uniform_pct > 0) asks that every class be sampled by the dataloader equally. In order to do this, we have to build a centroid file which describes, for each file, what classes are contained in a given image, and where the centroid for those classes is. If you have created your dataloader using the mapillary dataloader as a guide, then everything should work. The only thing i might suggest is changing class_uniform_tile to 860x480. The tile size means how many sets of distinct regions/hotspots to consider in a given image. In cityscapes, the tile size is 1024, which results in 2 tiles for 2k x 1k images.
GuessWhosBack
Raj-08
Hey, I was trying to adapt the code to train with my dataset. My dataset at the moment is rather small and contains only 6 classes. I wrote my own dataloader similar to the one from mapillary. It all works fine until I come across the centroid file. It works just running without it (just setting class_uniform_pct: 0). Thats why I wanted to ask if someone can explain the meaning of this file in a few sentences?
The apparent problem I ran into is that because of my dataset being small there a some classes where the centroid file has no centroids for: Found 4 centroids while it should be 6. Is there a way to adjust building this file to cover all classes? ( maybe setting class_uniform_til to something small like 256)
I have around 500 Pictures roughly 860x480 containing 6 classes.