Hi,
in the description it is mentioned to use the train validation split, "we use the YT-UGC dataset challenge split -- see Get Data & Get Data > Files.".
However, I'm not able to find the "Get Data > Files" part, nor am I able to find the description of the train validation split in the YT dataset.
Could you help me on this?
Furthermore, you have a requirement of 33ms for the processing of frames, is this measured for the per-frame calculation or as an average for the full video?
(e.g. it may take time to read the video, to move it to the GPU memory, and then to process it per frame, so I think there are several steps of processing time).
Best
Steve
Hello, you can find the train/test split in Participate > Files, just download the starter kit, there you have all the info :)
also I think you can download the starter kit from here: https://codalab.lisn.upsaclay.fr/my/datasets/download/a5da5970-47be-4998-bd57-a3bcb06b60f4
Best,
Marcos
You need to process at least 30 frames in one second (30FPS). If you sample those 30 frames to 5-6, and you can process them under 1s, it will work :)
We will feed 30 frames into the model and calculate the time it takes to process them, we will repeat that process 5 times and report the avg time.
Be creative in finding ways of reducing the time it takes to process 30 frames.
We will provide the evaluation framework in a few weeks.
Posted by: nanashi @ Feb. 14, 2024, 11:59 p.m.