Tensorflow - classification of videos

Question

Is the open-source version of TensorFlow suitable for classification of videos - akin the examples stated in http://vision.stanford.edu/pdf/karpathy14.pdf ? If not - are there any other packages in the public domain that could be used for this purpose?

Answer 1

It would be possible to modify the image models shipped with TensorFlow to train a neural network on such video data.

The main challenges would be:

1. Creating a custom reader that handled the appropriate video file format (or converting your video data to one of the supported format).

2. Using the image ops to generate the multi-resolution images required in the Karpathy et al. model.

3. (Possibly) using queues to perform pipelined video processing and ensure that I/O doesn't become a bottleneck.