ywchao / ho-rcnn

Code for reproducing the results in "Learning to Detect Human-Object Interactions"
66 stars 17 forks source link

Definition of rare classes #6

Closed baldassarreFe closed 4 years ago

baldassarreFe commented 4 years ago

I'm trying to reproduce some results on the HICO-DET dataset, but I don't seem to find the correct split between rare and non-rare classes.

According to the paper, rare hoi classes are those having <5 instances in the training set. Rare classes should be 167 of the 600 hoi classes in total.

However, if I filter the training set using this criterion I find these counts:

count<k    num_hois
 5              100
 6              114
 7              123
 8              129
 9              132
10              138
11              144
12              154
13              159
14              164
15              168
16              174
17              178
18              183
19              187
20              190

I checked your evaluation code and I spotted a <10 in eval_one.m, but even with that threshold, the number 167 doesn't come up.

Can you provide a list of exactly which hoi categories you consider rare and which not? A format similar to the one you have on your website would be ideal.

Thanks!

ywchao commented 4 years ago

The numbers you referred to ("<5 instances", "167 out of 600") should be for the HICO dataset (see ICCV'15). For HICO-DET (see WACV'18) the corresponding numbers should be "<10" and "138", which are exactly what you showed above.

baldassarreFe commented 4 years ago

Thanks for the quick reply! I must have missed the new definition of rare classes in HICO-DET, I assumed it was the same as HICO. It's all clear now ;)