improve generate_segments

[escorciav]

I wrote this function replicating the idea of the cvpr-paper. It would be great to refine it in the following aspects:

• [x] Apply intersection-threshold less than 1 just on annotations with length similar or greater that t_size.
• [ ] Segments with high coverage ratio should be more likely to be selected i.e. sampling based on cov_ratio_per_segment.
• [x] define appropriate values for RATIO_INTERVALS
• [x] main bug is here. index i is iterating over videos while score.shape[0] corresponds to the number of segments for all the videos. My fault :disappointed:

Note: if t_size is closer to l_size is possible that not enough segments are generated to satisfy a uniform distribution on RATIO_INTERVALS. We should decide how to overcome it.

[escorciav]

main bug was updated + notebook with example of how to use compute_priors. It would be great that you check if something looks weird there anyway :wink: