5faafede261a08fdaba46_source.mp4 -

A heavy network (like ResNet) extracts "deep features" only from select frames.

This allows for high-speed recognition because computing optical flow is much faster than running a full deep neural network on every single frame. 🛠️ Key Components 5faafede261a08fdaba46_source.mp4

For intermediate frames, the model uses a "flow field" (optical flow) to warp and move the previous features forward. A heavy network (like ResNet) extracts "deep features"