CinemaNet is a set of data and trained models to help run inference to classify images / frames of a video with an eye for photographic, cinemgraphic, composition and color labelling.
CinemaNet aims to give out of the box useful classification of images / frames of video to cinematographers, editors, archivists, and anyone interested in extracting classification in a cinema / video context.
The CinemaNet project aims to make a quasi knowledge graph of visual concepts useful to cinematographers, photographers, artists, designs, illustrators - and as such has labels ranging from composition theory to shot locations. The first round of label concents and categories is meant to provide an immediately helpful set of concepts and provide a baseline for the future. Note our label naming scheme uses a reverse DNS system - where top level naming helps to provide context for interpreting categories, concepts and sub-concepts.
See Labels.md for a list of categories and their concepts.
Note, the raw data set imahges are only useful if you plan on training your own models or are interested in helping optimize, classify and iterate on the quality of the models. Generally speaking its probably not needed!
Ensure that you have git checked out this repository or have done a download of this repository via the green clone or download button on the project page.
Ensure you have PIP installed. Install Google Image Downloader and install Google Chrome Driver with a matching version to your currently installed Google Chrome browser (for me, it was 74.x). Google Chrome Driver is required to download more than 100 images per google image query.
sudo easy_install pip in your Terminal.app command line.
pip install google_images_download in your terminal.app
Check that you have a version of Google Chrome installed in the defaul /Applications/Google Chrome.app location.
Launch Chrome and check the version number by going to 'About Chrome' in the Chrome Menu.
Download a matching version of Google Chrome Driver and place it into the same directory as these scripts.
You can then run python synopsis_categories_and_concepts_image_downloader.py to get the unfiltered raw data set - which will contain some noisy / misclassified images in the training set due to how Google Images returns results.
This download should be roughly 7.5 GB and contain roughly 63 thousand images sorted into a folder structure for the label category and concepts. The data set then requires manually pruneing from irrelevant or off topic images from the folder structure.
You can follow along with the Running Training Notes to see the steps we are taking if you want to train yourself.
Sign up for Googles AutoML Vision cloud service if you want to train your own model. At the time of this writing you will get approximately $300 in free credits.
See Running Training Notes for more info on training a model.
If you want to use the other utilities (auto labeler / HTML prediction preview script, model metadata and label name clean up script, and label print out script) you need to install Apples coremltools python package.
You can install via:
pip install -U coremltools