Reproduce “HPatches: A benchmark and evaluation of handcrafted and learned local descriptors” 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

Vassileios Balntas, Karel Lenc, Andrea Vedaldi, Krystian Mikolajczyk

The HPatches benchmark provides a comprehensive framework for evaluating local image descriptors, revealing that descriptor performance is highly task-dependent and that normalized handcrafted descriptors remain competitive with many learned features. It introduces a large, diverse patch-based dataset with defined evaluation protocols to address shortcomings of previous benchmarks.

Key insights

  • Descriptor performance and ranking are highly dependent on the specific evaluation task, highlighting the necessity of multi-task benchmarks.
  • Simple, unsupervised post-processing normalizations (e.g., ZCA, L2) significantly boost the performance of both handcrafted and learned descriptors, often making traditional methods competitive with or superior to deep learning approaches.
  • Patch verification alone is an insufficient metric for evaluating descriptor utility in real-world applications, as evidenced by the trivial MSTD descriptor performing well in verification but failing in other tasks.

Summary by alphaXiv.

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