Skip to content

UMAD-homo-eva-dataset.md

Latest commit

 

History

History
61 lines (37 loc) · 2.88 KB

UMAD-homo-eva-dataset.md

File metadata and controls

61 lines (37 loc) · 2.88 KB

UMAD Homography Evaluation Dataset(under maintenance)

UMAD-homo-eva dataset Download Link: Google Drive, about 290MB, includes the UMAD-homo-eva dataset overview and adaptive warping video, with 414 pairs of reference and query images, their pose data, and corresponding ground truth feature point annotations in a .txt files.

Introduction

The UMAD-homo-eva dataset project provides the following components:

UMAD-homo-eva dataset

dataset-overview

We chose 400+ image pairs from UMAD original dataset; For each evaluated image pair, we manually annotated 10 uniformly distributed matching points for quantitative comparisons by pre-labeling them using traditional feature point extraction and matching methods, use it to evaluate the image alignment algorithms.

UMAD-homo-eva dataset:

  • UMAD-homo-eva dataset overview and adaptive warping video,
  • 414 pairs of reference and query images,
  • 414 pose data: The pose data is collected by our robot system using 3D LiDAR SLAM/localization. We use the rotation data from the pose data for Rotation-induced Homography Warping (coarse) in our Adaptive Warping,
  • ground truth .txt files: 414 .txt files, corresponding to the image pairs.

Adaptive Warping

adaptive-warping

Adaptive Warping is a coarse-to-fine image alignment (or image warping) method based on homography.

PME-result

You can view more visualizations from the UMAD IROS 2024 video, or obtain more comparative data from the Leaderboard.

Leaderboard

Leaderboard-from-UMAD

We plan to add comparisons with more learning-based methods in the future.

Feature Correspondence Point Annotation Tool

We are optimizing this annotation tool. Currently, this annotation tool uses traditional feature point detection and matching methods (SIFT) for pre-matching, followed by manual screening, and finally allows for manual annotation.

Reference: https://github.com/daisatojp/labelMatch

Citation

This project is part of UMAD. If you find this work useful, please consider citing the paper:

@article{li2024umad
  author    = {Li, Dong and Chen, Lineng and Xu, Cheng-Zhong and Kong, Hui},
  title     = {UMAD: University of Macau Anomaly Detection Benchmark Dataset},
  journal   = {arXiv preprint arXiv:2408.12527},
  year      = {2024},
}