jinwchoi / computer_vision_literature_review Goto Github PK

Hi Choi,

Firstly, thank you for your work on collecting recent paper on action recognition.

Second, I have a question and I hope we can discuss it here. In almost all recent paper on action recognition topic, they resized the original frame image from (320*240*3) into (300*300*3) or other size, such as in [1], [2], [3].
That makes me wonder what about the ground truth? The ground truth which is 4 number of pixel coordinates, is originally for original video dataset (320*240*3) and when they resize the image, did they resized the ground truth into (3003003) as well or use the original one? Because their algorithms usually based on the ground truth overlapping for evaluating the accuracy, so I think it will make a big difference in the result.
By the way, do you know which version for the ground truth they used in their work for UCF-101 24 classes and other datasets? I am asking because I found UCF-101 has 2 version of the ground truth: one is the original and the other one is remade by Gurkirt et al. you can find the link here in the Additional Data section. This gets me confuse.

Thank you