I have the video I want to analyse and the frame and annotation for the training separately, I was thinking if it would be possible to use Osvos directly with the video, without extracting its frames in order to reduce the space used and to improve (possibly) the performance. I suppose that the changes must be done in the class dataset in the initialization, get_text_size and next batch functions.
I have tried converting the pointers parts of the next_batch function phase test into reading the video for X frames, but the result has been always wrong. I do not know if there is a method implemented either in this implementation or in the Caffe or pytorch ones.
I have the video I want to analyse and the frame and annotation for the training separately, I was thinking if it would be possible to use Osvos directly with the video, without extracting its frames in order to reduce the space used and to improve (possibly) the performance. I suppose that the changes must be done in the class dataset in the initialization, get_text_size and next batch functions. I have tried converting the pointers parts of the next_batch function phase test into reading the video for X frames, but the result has been always wrong. I do not know if there is a method implemented either in this implementation or in the Caffe or pytorch ones.