A question about v-coco dataset

[ltttpku]

Hi, thank you for your code; But i have a question about the V-COCO dataset you implemented.

V-COCO is a subset of COCO dataset and has 10, 396 images (5,400 for training and 4,964 for testing) as stated in paper.
However, when executing your code for v-coco dataset, I foundlen(trainset)==4969and len(testset)==4532. The reported number of training images does not match the actual one.

[fredzzhang]

Hi @ltttpku,

If I remember this correctly, there are images without sufficient bounding box annotations. So I removed those images. It's the same situation with HICO-DET as well. The authors claimed there are 38118 training images, while in fact only 37633 of those images have bounding box annotations.

Fred.

[ltttpku]

ohh, I see. Thanks!

[hutuo1213]

Hi @fredzzhang,

I found 5400-431=4969 valid actions in instances_vcoco_trainval.json. This is different from the 4971 published by PPDM, what causes this discrepancy? How are your annotation files collected? Hope to get your guidance!

yaoyaosanqi.

[fredzzhang]

Hi @yaoyaosanqi,

The files instances_vcoco_*.json were converted from the original annotation files released by Gupta et al. I'm not sure what the authors of PPDM did with their annotations, so I do not know why there is a discrepancy. It's possible that they didn't remove images for the verb point, which do not contain valid HOI triplets.

Fred.

[hutuo1213]

We looked at and ran the code, as you said.
Cheers!