Reproduce “Resource Efficient 3D Convolutional Neural Networks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Okan Köpüklü, Neslihan Kose, Ahmet Gunduz, Gerhard Rigoll
Researchers from TU Munich and Intel Labs developed and evaluated resource-efficient 3D Convolutional Neural Networks, adapting common 2D architectures for video processing, demonstrating competitive accuracy with significantly reduced computational cost and real-time performance on edge devices. Their 3D-ShuffleNetV1 2.0x model achieved accuracy comparable to ResNet-18 with 7 times fewer parameters and 15 times fewer FLOPs.
Key insights
- Effective feature learning in efficient 3D CNNs necessitates sufficient network depth; excessively shallow or narrow designs (width multiplier < 0.5) significantly degrade accuracy.
- Depthwise convolutions, as seen in 3D-MobileNetV2, are particularly adept at capturing temporal motion patterns, proving highly effective for motion-centric tasks like gesture recognition.
- Theoretical FLOP counts are often insufficient predictors of real-world runtime performance, as memory access costs, parallelism, and hardware-specific optimizations play a crucial role.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.