Reproduce “LP-3DCNN: Unveiling Local Phase in 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

Sudhakar Kumawat, Shanmuganathan Raman

Researchers from the Indian Institute of Technology Gandhinagar introduced the Rectified Local Phase Volume (ReLPV) block for 3D Convolutional Neural Networks, efficiently extracting local phase information to significantly reduce model parameters and computational expense. The method achieved top performance on 3D object recognition and video action recognition benchmarks, outperforming prior state-of-the-art with substantially fewer parameters.

Key insights

  • Local phase information, derived via STFT, is highly effective for representing perceptually significant features like edges and contours in 3D data, analogous to traditional convolutional filters.
  • The STFT inherently decorrelates input signals, acting as a natural regularizer that reduces overfitting and improves generalization in deep 3D CNNs.
  • The ReLPV block drastically reduces the number of trainable parameters (e.g., 27 times for a 3x3x3 filter) and computational costs compared to standard 3D convolutions, offering substantial efficiency gains without performance degradation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.