No action here just another note to self. Much of difficulty with training deep models is due to low instances of videos. I created another view of the data that splits videos into train/test/val, splits each video into chunks of frames (can sample with gaps between frames), and then shuffles these around within their train/test/val bins. Goal is to model the processes in each chunk. This split takes advantage of the breadth of data, but much of the data samples in training will look similar to each other, since taken from the same video. Not sure how much this will effect generalizing to new videos.
Just keep in mind good practices for the thesis - take advantage of any pretraining on similar/synthetic data to combat small dataset. Pay plenty of attention to regularization techniques to stave off brittle features learn from low variance in train set.
No action here just another note to self. Much of difficulty with training deep models is due to low instances of videos. I created another view of the data that splits videos into train/test/val, splits each video into chunks of frames (can sample with gaps between frames), and then shuffles these around within their train/test/val bins. Goal is to model the processes in each chunk. This split takes advantage of the breadth of data, but much of the data samples in training will look similar to each other, since taken from the same video. Not sure how much this will effect generalizing to new videos.
Just keep in mind good practices for the thesis - take advantage of any pretraining on similar/synthetic data to combat small dataset. Pay plenty of attention to regularization techniques to stave off brittle features learn from low variance in train set.