This toolbox can be used to train a model using two sets of images, and then sort validation/target images; rename images to assist in time lapse creation; overlay a map on geotagged time lapse images.
- Build model: Run generate_classifier.py on a directory of desirable (known as good/positive) and undesirable (known as bad/negative) images. This outputs a sklearn.model_selection.GridSearchCV model.
- Execute model: Run sort_images.py on a directory of images to sort them based upon the model built in step 1.
Additionally, the following utilities are included to facilitate time lapse video creation:
- Re-number files: Run batch_rename.py to sort and renumber (
--renumber) all images files, padded with zeros to match the maximum file count. The software ffmpeg requires input images to be numbered as such. - Pull specific time(s) of day: Run daily_subset_and_rename.py to grab images between specific hour(s) of day, move to new folder, and optionally call batch_rename.py to re-number the files after they are moved.
- Decimate files: Run reduce_frames.py to remove files based upon a "keep factor" (e.g., a factor of '4' keeps every fourth image.) Images will automatically be renumbered, but can be disabled.
Run add_map_to_timelapse.py to add map to images (note: only works with GoPro's geotags.)
I have written an example of how to use these tools for filtering unlit images captured by a time-lapse camera on my blog.
- numpy
- scikit-learn
- pillow (PIL)
Conda one-liner to create environment:
conda create --yes --name timelapse-tools -c conda-forge python=3 numpy scipy scikit-learn pillow
I credit parts of the image classification work to this blog post: http://www.ippatsuman.com/2014/08/13/day-and-night-an-image-classifier-with-scikit-learn/