Purity Eval - Bounding Box Crop?

[ThomasNorr]

Dear Authors,

thank you for this fascinating work.

I am wondering about the Purity Evaluation. As you mentioned, you used the cropped images following ProtoTree (https://github.com/M-Nauta/ProtoTree/blob/main/README.md#preprocessing-cub). However, I am not seeing how you adapt the part annotations according to these crops? Am I missing that ?

Best Regards,
Thomas

[ThomasNorr]

As a followup, I have seen that you used the cropped images for evaluting purity too. However, so far I was unable to produce the purity results reported in the paper for the published Resnet model, while accuracy matches. Does this model reach the reported ~60% with your setup?

Also, the figure you report in table 3 comes from these "topk" lines (

PIPNet/main.py

Line 308 in 12c4da7

csvfile_topk = get_topk_cub(net, projectloader, 10, 'train_'+str(epoch), device, args)

), right?

    csvfile_topk = get_topk_cub(net, projectloader, 10, 'train_'+str(epoch), device, args)
    eval_prototypes_cub_parts_csv(csvfile_topk, parts_loc_path, parts_name_path, imgs_id_path, 'train_topk_'+str(epoch), args, log)