HPatches

Dataset Overview

Dataset Name

HPatches: Homography‑patches dataset

Dataset Purpose

Used to evaluate the performance of local descriptors, especially under illumination and viewpoint variations.

Dataset Structure

• Sequence Types:
  • i_X: Illumination‑change sequences
  • v_X: Viewpoint‑change sequences
• File Formats:
  • ref.png: Reference patches from the reference image
  • eX.png and hX.png: Corresponding patches from other images, where e denotes easy and h denotes hard
  • Each patch is 65 × 65 pixels and stored in a single .png file, with patches stacked in a single column

Patch Extraction Method

• Reference and Target Images:
  • Each sequence contains one reference image and five target images that differ in illumination and/or viewpoint.
  • Ground‑truth homography (H) is provided for each target image relative to the reference.
• Patch Extraction:
  • Patches are sampled in the reference image using local feature detectors (Hessian, Harris, and DoG).
  • Patch orientation is estimated by Lowe's method; no affine adaptation is applied; patches are square.
  • Patches are extracted from a region 5× larger than the detected feature scale to ensure full coverage.
  • To avoid duplicate detections, detections with >50 % elliptical overlap are clustered and a random one is retained.

Dataset Download

• Patch dataset:
  • HPatches [4.2 GB]
• Full image sequences:
  • HPatches full sequences [1.3 GB]

Citation

• Reference:
  • HPatches: A benchmark and evaluation of handcrafted and learned local descriptors, Vassileios Balntas*, Karel Lenc*, Andrea Vedaldi and Krystian Mikolajczyk, CVPR 2017.