Mar 2021 · Model Compression / Temporal Redundancy / 58M FLOPs / Edge AI / Video Understanding
PatchNet: Ultra-Compact CNN for Edge Computing
An ultra-compact CNN that reuses temporal information in video to reduce computation to 58M FLOPs. It achieves 5× fewer operations and 4.3× faster inference on edge devices at the same accuracy. The research and implementation are publicly available.
The challenge
Video inference repeatedly processes related frames. Reusing this information offers a way to reduce computation while preserving recognition accuracy.
The approach
PatchNet exploits temporal redundancy with an ultra-compact CNN design. The public paper and code describe the model and its implementation.
The outcome
58M FLOPs, with 5× fewer operations and 4.3× faster edge inference at the same accuracy.