Continuous online video classifcation
This repo holds the code that supports our blog posts about using TensorFlow, Inception and a Raspberry Pi for continuous online video classification.
Part 1 - CNN, Inception only: https://medium.com/@harvitronix/continuous-online-video-classification-with-tensorflow-inception-and-a-raspberry-pi-785c8b1e13e1
Part 2 - Adding an RNN (LSTM): https://medium.com/@harvitronix/continuous-video-classification-with-tensorflow-inception-and-recurrent-nets-250ba9ff6b85#.3vl3apzb6
Overview
Step 1:
Capture a bunch of video with stream_images.py
.
Step 2:
Move each frame into its class directory in images/classifications/[classname]/ - you can do this automatically by defining the start and stop timestamps of each commercial in the commercials.py
script, and then running build_labels.py
. Be sure to set copyimage to True if you want it to copy the image. This also creates a reference file that we use later that identifies the class of each image.
Step 3:
Run the tensorflow/examples/image_retraining/retrain.py
script in the main TensorFlow repo. The full command we use is in the blog post linked above.
Step 4:
Run make_predictions.py
on the holdout set to see how it does.
Step 5:
Run the online system with online.py
on your Raspberry Pi, which will classify each frame captured with our newly trained weights.