objects prediction in script/extract_features.py

[BigodF]

Following code for predicting object in script/extract_features.py

# Predict the class label using the scores
objects = torch.argmax(scores[keep_boxes][start_index:], dim=1)
cls_prob = torch.max(scores[keep_boxes][start_index:], dim=1)

my understanding is as following:
the size of scores[keep_boxes] is (100, 1601), start_index is for classes, prediction objects exclude foreground when it is 1.
the size of scores[keep_boxes][start_index:] is (99, 1601), what we want is (100, 1601) or (100, 1600) for prediction scores.
so the code should be writen like this if we want to compare the prediction scores of expected classes:

# Predict the class label using the scores
objects = torch.argmax(scores[keep_boxes][:, start_index:], dim=1)
cls_prob = torch.max(scores[keep_boxes][:, start_index:], dim=1)

[prajwalgatti]

Yes, this is a bug.

But doesn't affect the current pre-training/fine-tuning procedure as the objects and cls_prob data is discarded when converting to lmdb formatted data (see conversion code)

But I agree this should be corrected, regardless.