How to get the small scale of data about self-supervised-temporal-discriminative-representation-learning-for-video-action-recognition HOT 2 CLOSED

It is a good question.

Both UCF/HMDB contain thousands level video, to explore the influence of more data. We sample different amount samples in Kinetics in class-balanced way as previous method like S3D. What's need is to define the sample radio and generate train list. Please look gen_sub_dataset.py.

from self-supervised-temporal-discriminative-representation-learning-for-video-action-recognition.

Hi, @FingerRec . Thank you for your instance response.

It seems I misunderstand some parts of this part and I thought it is something like pretraining on large dataset and finetuning on small scale data (which is something like semi-supervised learning in SimCLRv2). I did not look carefully about that part especially the figure because UCF101 does not contain more than 15,000 videos. Sorry for that.

Anyway, you have solved my problem and I will follow your instructions. Thank you.

from self-supervised-temporal-discriminative-representation-learning-for-video-action-recognition.