search

genbing99 / SoftNet-SpotME

Shallow Optical Flow Three-Stream CNN For Macro and Micro-Expression Spotting From Long Videos
44 stars 20 forks source link

How to choose the best model parameters? #5

Closed ADstudyer closed 2 years ago

ADstudyer commented 2 years ago

May I know how to choose the best model parameters? or you just save the model after training 10 epochs?

genbing99 commented 2 years ago

The model parameters are saved after training 10 epochs.

ADstudyer commented 2 years ago

Hello, I have the following questions:

  1. Why move the ground truth forward by K frames in the spotting process? In training.py line159: for samples in video: gt.append([samples[0]-k, 0, samples[1]-k, 0, 0, 0, 0]) total_gt += 1

  2. When applying for the SAMM long videos dataset, my application email was regarded as spam by the server and could not be sent to the mailbox of the person in charge. I'm curious how did you apply?

genbing99 commented 2 years ago
  1. The adjustment is due to the score aggregation mentioned in the paper Section 2.4 where we shifted the spotted intervals forward by k frames, therefore the ground-truth labels are required to be shifted.
  2. I am not sure about this, you may email the authors for more details.