achalddave / human-labeling-videos

Image labeler

Example: Labeling ImageNet Video Frames

To label some sample images, first edit the IMAGES_ROOT variable in imagenetvid.cfg to point to the root of your ImageNet Vid dataset.

FLASK_DEBUG=1 FLASK_APP=label.py FLASK_CONFIG=imagenetvid.cfg \
flask run -p 8000 -h 0.0.0.0

IGNORE BELOW FOR NOW; instructions below not updated from original video labeler used by @achald for a different project.

Filtering

Once videos are labeled, you can filter the results by using the filter_labels.py script, like so:

python filter_labels.py \
    --input-labels <path-to-labels> \
    --must-have ${MUST_HAVE_LABELS} \
    --must-not-have ${MUST_NOT_HAVE_LABELS} \
    --output-labels <path-to-filtered-labels> \
    --labels-list <path-to-labels-list>

In general, the you should use the following setup:

perfect - moving-unlabeled - MUST NOT HAVE static-labeled - MUST NOT HAVE no-moving-labeled - MUST NOT HAVE oversegmented-moving - MUST NOT HAVE static-part-labeled - MUST NOT HAVE interesting-video - look-again - MUST NOT HAVE challenging - always-moving -

In a sense, the 'perfect' label is kind of redundant; it just implies that moving-unlabeled, static-labeled, and no-moving-labeled are all false.