Which Deep Learning architecture is best for classifying short videos of variable length? I would like to classify videos that last from 1 up to 3 seconds.
My suggestion is to use Convolutional and Recurrent layers in the same Neural Network.
You'd have to capture a given number of frames of a video (let's say one each 0.5 seconds), and feed arrays of screenshots into the model. Its structure would be:
- Conv (and MaxPool) layers to process pixel data - they will extract and process relevant information from each screenshot.
- LSTM layers - that will process their sequence, extracting meaning from their flow.
- Dense layers at the end to perform classification, with softmax activation at the output layer.
That's how I would do. It's going to be computationally expensive, if you don't have a GPU it won't be easy.