the sizes of different dataset

[tian-2024]

Hi, I would like to verify if the dataset sizes I printed here are correct.

It seems that for Human3.6M, the aligned and unaligned datasets have different paths. My assumption is that the aligned dataset is the one previously downloaded, while the unaligned dataset uses the original dataset with some images filtered out.

Dataset	Train	Test
taichi	5000	300
cub_aligned	5964	2874
cub_001	29	17
cub_002	30	14
cub_003	30	15
cub_all	5964	2874
deepfashion	10604	1179
celeba_aligned	19000	1000
celeba_wild	5379	283
human3.6m	796648	87975
unaligned_human3.6m	159444	17615

Additionally, since the evaluation batch size is set to 1, would this process be very time-consuming, especially for the Human3.6M dataset?

The below are time records of my experiments:

Dataset	Eval Time	Images
taichi	0.1 hours	300
celeba	0.5 hours	1000
deepfashion	1 hour	1179
cub	2 hours	2874
human36m	-	87975

I guess human36 evaluating will be 60 hours?

[ehedlin]

Yes, the evaluation for human3.6m was definitely time consuming. And for the unaligned human3.6m dataset I used this repository to install/process the dataset. This discards frames that are too similar.

And I dont have the datasets downloaded anymore but those lengths look right to me.

[tian-2024]

Yes, the evaluation for human3.6m was definitely time consuming. And for the unaligned human3.6m dataset I used this repository to install/process the dataset. This discards frames that are too similar.

And I dont have the datasets downloaded anymore but those lengths look right to me.

Thank you so much!! I tried to find the file for long time and failed, your advice is very useful and important~