Action-Recognition

Optimized Action Recognition to enable faster inferences using just images instead of videos.

Drawbacks of Current Techniques:

• Action Recognition has been a compute heavy process where a video is analysed frame by frame using a Recurrent Neural Network and then have their results across time combined to detect the action performed.
• Apart from RNNs making the process compute heavy, it is also time consuming for a larger set of videos.

Ideas for improvement:

• Can a video be converted to a smaller format?
• Is there another way to extract temporal information?

Solutions to these issues:

• Eliminate use of videos by bringing in images
• Utilize Optical Flow

The notebooks contain snippets on how the data was processed and sent to the model and how well the model learned and performed. Further improvements include:

• Improving frame selection
• Training on large data using Unsupervised Learning
• Supporting newer model architectures