Reproduce “Parameter-Efficient Fine-Tuning in Spectral Domain for Point Cloud Learning” 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
Dingkang Liang, Tianrui Feng, Xin Zhou, Yumeng Zhang, Zhikang Zou, Xiang Bai
PointGST introduces a parameter-efficient fine-tuning method for point cloud learning that operates in the spectral domain, achieving state-of-the-art performance on various 3D tasks with a small fraction of trainable parameters. It notably reaches 99.48% accuracy on ScanObjectNN OBJ BG, the first method to exceed 99% on this benchmark, using only 0.67% of the parameters of a fully fine-tuned large model.
Key insights
- Spectral representations derived from graph theory inherently de-correlate confusing features into orthogonal frequency components, simplifying optimization for task-specific adaptation.
- Incorporating intrinsic geometric information of the downstream point cloud directly into the tuning process through data-dependent graph construction and spectral bases enhances adaptation effectiveness.
- Multi-scale spectral bases (global and local) allow the model to capture both broad and fine-grained geometric structures, leading to more robust and comprehensive feature adaptation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.