How can I use this model in micro-expression real-world scenarios without labeled onset frames?

[earliesttttt]

As mentioned in the paper, the D&C-RoIs technique can be used to find the peak frames. However, in real-world micro-expression scenarios, there are no personnel to help label the onset frame positions. Therefore, we do not know when the micro-expression begins, making it impossible to generate optical flow maps as described in the paper. How should we handle the video data in this case?

[huazai6161]

It's called micro-expression spotting, saperated from recognition (classification).