Simple command line video processing using Tensorflow and TF-SLIM.
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Type Name Latest commit message Commit time
Failed to load latest commit information.


Using Tensorflow & TF-SLIM to Detect Features in Video

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 (
  • 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
$ cd tf-video
$ python -h


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

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

Complex example: --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]

Additional Switches & Options

--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