BBoxes from TREK-150: confused as to format about trek-150-toolkit

[matteo-dunnhofer]

Hi @vineetparikh

I noticed that you are decoding frames with VideoReader. We used ffmpeg to decode frames. I suspect that the two methods decode the frames differently (we experienced something similar in other projects), and so there might be a different number of frames per video, ultimately leading to a wrong synchronisation of frames and bounding boxes. Can you try to run your code on the frames we release in TREK-150? You can download them using the toolkit.

[vineetparikh]

^ will do and will report back if this is the case!

[vineetparikh]

looks like different ffmpeg versions were the case (decord packages a different ffmpeg version). Pulling the dataset via the toolkit, I'm getting more reasonable bboxes for the same objects at least for this sequence, so the frame IDs need to be extracted in exactly the same way.

Thanks for the clarification! I'm going to close this issue but hopefully this will be helpful to someone else who has a similar problem moving forward.