Generate time frames from time point annotations in scene-recognition

[marcverhagen]

Because

No response

Done when

No response

Additional context

No response

[keighrim]

Current proc.py in SR project only re-formats point-wise annotations into (slightly simpler) point-wise annotation CSV as gold. (#83) To address this issue, we need a few changes in the structure of this repo so that we can

1. allow subdirectories under golds to hold different "kinds" of gold annotations OR
2. proc.py (in the new credits time frame directory) to read "raw" files across annotation projects

(issue 2 can be easily addressed by using symlink, I think)

[marcverhagen]

For now I am working with timepoints and timeframes subdirectories.

Not sure what item 2 above refers to, but if it is what I think it is then this is a more general issue and applies to any two annotation projects with the same data structures.

[marcverhagen]

Done in 9ae2e00 and a6104fe

[marcverhagen]

Reopening because not all issues in the comment from @keighrim were addressed

[keighrim]

I think the (non-)description of this issue has become somewhat outdated after we received SR annotations¹ that's not done in "sequential" way, and hence can't really be converted to timeframe annotations. Given those new annotation, the approach with golds/timepoints and golds/timeframes subdirectories (suggested in #102) would bear some problems moving forward;

1. timeframe intervals in csv files generated from the non-sequential annotations do not necessarily cover their underlying scene intervals
2. golds/timeframes subdirectory holds heterogeneous data - some naturally sequential annotation that (approximately) cover entire length of videos, other fractional interval annotations².

If the evaluation script that addresses clamsproject/aapb-evaluations#43 (currently in development in https://github.com/clamsproject/aapb-evaluations/tree/43-swt-eval) needs timeframe-based gold data, we might want to consider either

• separating timeframe golds into something like scene-recognition-timeframe as a derivative from scene-recognition and proc.py there only handles sequentially annotated batches.
• handling the conversion in the evaluate.py script - this might involve more work both in coding and documenting.

Footnotes

1. https://github.com/clamsproject/aapb-collaboration/issues/27#issuecomment-2411942010 ↩

2. This might not be a big deal since scripted evaluations will always use batch information and won't necessarily use all the files in a gold directory. ↩