An automation "wrapper" based on TF-SLIM to make it easy to detect various features in video using Tensorflow, FFMPEG, and various versions of the Inception neural network.

• Can utilize virtually any trained model (Inception Resnet V2, Mobilenet, etc) via Tensorflow Hub (retrain.py).
• Runs under both Linux and Windows (Python 3.6.x).
• Silly fast, even on CPU for detection. Using an GTX 1060 Nvidia GPU it can perform analysis of a two-hour video in less than 7 minutes even using Inception Resnet V2 and under 3 minutes using Mobilenet V1.

Tensorflow 1.x (tested on 1.7), FFMPEG (Tested on 3.4), and Python 3.6.x (Windows & Linux). A pre-trained model such as Inception V3 or Mobilenet.

No formal "installation" is required beyond making a copy of this directory on your local system.

$ git clone git@github.com:robertrittmuller/tf-video.git
$ cd tf-video
$ python videoscan.py -h

Search video for features in video. Creates/overwrites [videofilename]-results.csv.

videoscan.py --modelpath models/mymodel.pb --labelpath models\mylabelsfilename.txt --reportpath ..\example-reports --labelname mylabel [myvideofile.avi]

videoscan.py --modelpath models/mymodel.pb --labelpath models\mylabelsfilename.txt --reportpath ..\example-reports --labelname mylabel --fps 5 --allfiles --outputclips --smoothing 2 --training --videopath [/path/to/video/files]

--modelpath Path to the tensorflow protobuf model file.
--modeltype Type of Tensorflow model being loaded (mobilenetV1, inception_resnet_v2, etc).
--labelpath Path to the tensorflow model labels file.
--labelname Name of primary label, used to trigger secondary model (if needed).
--reportpath Path to the directory where reports/output are stored.
--temppath Path to the directory where temporary files are stored.
--trainingpath Path to the directory where detected frames for retraining are stored.
--height Height of the image frame to be extracted and processed. Needs to match model input layer!
--width Width of the image frame to be extracted and processed. Needs to match model input layer!
--smoothing Apply a type of "smoothing" factor to detection results. (Range = 1-6)
--fps Frames Per Second used to sample input video. The higher this number the slower analysis will go. (Default is 1 FPS)
--allfiles Process all video files in the directory path.
--outputclips Output results as video clips containing searched for labelname.
--training Saves predicted frames for future model retraining.
--outputpadding Number of seconds added to the start and end of created video clips.
--filter Value used to filter on a label. [Depricated]
--keeptemp Keep temporary extracted video frames stored in path specified by --temppath

MIT