Extracted features does not match the split in COCO dataset

[seanzhuh]

Hi, I've downloaded the pretained features for train/val using the provided links (36 - 100 features per image), however, there are four tsv files in it, namely,

1. karpathy_train_resnet101_faster_rcnn_genome.tsv.0
2. karpathy_train_resnet101_faster_rcnn_genome.tsv.1
3. karpathy_val_resnet101_faster_rcnn_genome.tsv
4. karpathy_test_resnet101_faster_rcnn_genome.tsv

each of the four file contains 56643,56644,5000,5000 unique image_id, however, in vqa2.0, training dataset has 82783 image, validation set contains 40,504 images, how can I know which image_id is for training set, which is for validation set?

[nihirv]

Struggling with this too. I've loaded in the trainval files and the test files but am still struggling to find images which were in the original 2014 validation set

[haixiongli]

Hello, I encountered the same problem. I don't know how the csv file here corresponds to the image id in the coco dataset. Have you solved this problem?