image similarity-based annotation for generating bootstrap data

[keighrim]

As a subtask of #4 , we'd like to experiment with the idea of having quick-and-dirty annotation environment for creating a relatively small bootstrap annotation, based on pseudo-clustering based on image similarity. Some rough sketch on the implementation.

1. run shot detection algorithm (e.g., https://www.scenedetect.com/, https://github.com/clamsproject/app-pyscenedetect-wrapper) on a set of video (or a fixed total duration of videos)
2. sample N frames from each "shot"
3. apply a feature extraction on all images
4. taking the first image (chronologically) as the reference, sort all images based on similarity to the reference image
5. present first N similar images to annotators and ask for a label, and membership verification
6. sort the leftover (unlabeled) images chronologically and go back to 4, until the number of remaining images drops under the threshold
7. go to the next batch of videos and start from 1, merging new feature vectors with the leftover