Compass Detection Algorithm based on Image Corner

Jia Li, Jinliang Wang, Ping DuanInternational Journal of Signal Processing Image Processing and Pattern Recognition 2013 · DOI

TL;DR

This paper introduces a novel "Compass Detection Algorithm" (ATCD) to improve upon the FAST corner detection method. ATCD addresses FAST's limitations in adaptive thresholding and handling of skewed edges, resulting in more accurate, well-distributed, and robust corner detection with comparable efficiency.

What Problem Does This Paper Address?

Corner detection is a fundamental step in many computer vision tasks like tracking, localization, and image matching. Existing methods, including the popular FAST algorithm, face challenges. FAST, while efficient, struggles with setting an appropriate threshold for corner detection, which is highly dependent on image properties. This can lead to either detecting too many false positives (pseudo-corners) or missing crucial corners. Furthermore, FAST can incorrectly identify points on skewed edges as corners, reducing accuracy. The paper aims to overcome these limitations by proposing an improved corner detection algorithm that is more robust, accurate, and adaptive to varying image conditions.

Methods

The proposed Compass Detection Algorithm (ATCD) builds upon the FAST algorithm but introduces several key improvements:

  • Adaptive Thresholding: Instead of a fixed threshold, ATCD calculates an adaptive threshold based on the image's absolute contrast (difference between maximum and minimum intensities). This allows the algorithm to better adjust to different image lighting and intensity distributions. The threshold is set as a percentage (15-30%) of this absolute contrast.
  • Compass-Based Pre-detection: ATCD uses four "compasses" as test masks. These compasses are formed by specific pixel arrangements around a candidate point. This approach is used for an initial corner pre-detection.
  • Skewed Edge Pseudo-Corner Elimination: A specific test is implemented to distinguish true corners from pseudo-corners that arise on skewed edges. This test examines intensity differences between opposing pixels on the circle around the candidate point.
  • Non-Maximum Suppression using Laplacian: To handle multiple adjacent responses and ensure only the most significant corners are kept, the algorithm employs a non-maximum suppression technique. This involves approximating the Laplacian of Gaussian (LOG) and selecting points with extremal Laplacian values.

Datasets

The paper does not mention any newly introduced benchmarks. The experiments were conducted using "Mikolajczyk05 [16] standard images set" for evaluating robustness under various transformations (rotation, viewpoint, illumination, scale change). Specific image sizes used for efficiency testing include 320×278, 480×640, 694×1066, and 2056×2452 pixels.

Key Insights and Ideas

  • Adaptive Thresholding is Crucial: The paper strongly emphasizes that a fixed threshold is a significant bottleneck for corner detectors like FAST, and an adaptive approach based on image contrast significantly improves robustness.
  • Compass Masks Offer a Structured Approach: Using predefined compass patterns for pre-detection provides a structured way to analyze local image structure, potentially leading to more reliable initial corner candidates.
  • Targeted Skewed Edge Handling: Explicitly addressing and eliminating pseudo-corners on skewed edges is vital for improving the accuracy and reliability of corner detection in real-world scenarios.
  • Laplacian for Refinement: Employing Laplacian for non-maximum suppression is a standard but effective technique for ensuring that only the most prominent corners are retained, leading to better distribution.

Key Results

  • Improved Accuracy and Distribution: ATCD demonstrates more accurately located and better-distributed corners compared to FAST, especially in eliminating pseudo-corners on skewed edges.
  • Robustness: Experiments show ATCD achieves good performance under illumination and scale changes, with superior repetition rates compared to FAST in some cases. While viewpoint transformation poses challenges for both, ATCD still shows competitive results.
  • Efficiency: ATCD maintains a comparable running time to FAST, indicating that the improvements in accuracy and robustness do not come at a significant computational cost. For instance, for a 694x1066 image, FAST takes 51.96ms while ATCD takes 53.47ms.

Limitations and Future Work

  • Limitations:
    • The paper notes that for scale change, both ATCD and FAST have low repetition rates, suggesting that neither algorithm is inherently designed to handle significant scale variations robustly.
  • Viewpoint transformation also presents challenges for both algorithms, indicating a need for further improvements in handling perspective changes.
  • Future Work: The paper doesn't explicitly detail future work, but the findings suggest that further research could focus on enhancing robustness to scale and viewpoint changes, potentially by incorporating scale-invariant features or more advanced geometric transformations.

Summary updated

Abstract

To improve the Features from Accelerated Segment Test (FAST) algorithm such as artificial threshold setting and the choice and ordering of continuous points on circle, a novel compass detection algorithm is proposed.The algorithm uses four compasses as the test mask and calculates threshold value based on image intensity feature.Four compasses are used to pre-detect corners.In order to improve the location accuracy of corners, pseudo response on skewed edges is eliminated.Finally, multiple adjacent responses are tackled by keeping only points which have extremal value of Laplacian.Comparative experiments between FAST and the proposed algorithm are carried out with respect to validity, robustness and efficiency.The results show that the detected corners by the proposed algorithm are accurately located and well-distributed.
BibTeX
@article{li2013compass,
  author = {Jia Li and Jinliang Wang and Ping Duan},
  title = {{Compass Detection Algorithm based on Image Corner}},
  journal = {International Journal of Signal Processing Image Processing and Pattern Recognition},
  year = {2013},
  doi = {10.14257/ijsip.2016.9.5.12},
}
Figure