A few questions about enriching the learning database

Hi,
I have a few questions about enriching the learning database for the same experiment:

  • is it possible to use a list (text file) for the path to the videos with the add_new_videos function?
  • after this add_new_videos step the extract_frames function does not work because the folders to add the extraction frames do not exist, I have to manually create each directory before launching extract_frame.

a more general question is: to strengthen the network’s capacities, I regularly add the images extracted from the last experiences: should I re-train the network (using last snapshot) with all the images marked since the creation of the project or only with the latest images added?

  • is it possible to use a list (text file) for the path to the videos with the add_new_videos function?

– no, currently you need to type in the path or folder, or use the GUI to select the videos. Placing the videos you want to use for labeling into one folder does work though…

  • after this add_new_videos step the extract_frames function does not work because the folders to add the extraction frames do not exist, I have to manually create each directory before launching extract_frame.

– Extract frames will work; you need to use an automatic method first; the idea is you want to use clustering techniques to extract maximally different frames. If you want to supplement with a few manually selected options after that, you can run the function again with ‘manual’ set.

“with all the images marked since the creation of the project or only with the latest images added?”

you must absolutely use all the data to retrain, not just the new images.

Thanks !

sincerely
gilles

1 Like

The config.yaml files can be easily manipulated and you can also write a targeted function to add multiple videos etc. Just check out how the source code does it for one example: https://github.com/AlexEMG/DeepLabCut/blob/2a2ad95a1ec35ad7f98d0155b4aa0d3ecedaefac/deeplabcut/create_project/add.py#L99

1 Like