ChESS - Quick and robust detection of chess-board features

Stuart Bennett, Joan LasenbyCVIU 2014 · DOI

TL;DR

This paper introduces ChESS, a novel feature detector for chess-board vertices that is fast, robust to noise and lighting, and doesn't require prior knowledge of the pattern's scale or orientation. It achieves this by using a unique sampling strategy and a response calculation that provides a strength measure, making it suitable for real-time applications like camera calibration and 3D reconstruction.

What Problem Does This Paper Address?

Accurately detecting chess-board vertices is crucial for many computer vision tasks, especially camera calibration and 3D structured light reconstruction. Existing methods often struggle with variations in lighting, contrast, and perspective, requiring manual parameter tuning and limiting their use in automated systems. While some detectors offer robustness, they can be computationally expensive or lack the precision needed for accurate calibration. This paper aims to fill this gap by providing a detector that is simultaneously fast, robust, and accurate, offering a strength measure for detected features.

Methods

The core of the ChESS detector lies in its novel sampling strategy and response calculation. Instead of relying on simple thresholding or general-purpose corner detectors, ChESS uses a circular sampling ring around a potential vertex.

  • Sampling Strategy: A ring of 16 pixels is sampled at equal angular intervals around a potential feature center. This sampling is designed to be rotationally invariant and to mitigate issues caused by optical blur and pixel quantization, which can occur at the edges of chess-board squares. The radius of this ring is empirically determined to be around 5 pixels for VGA resolution images.
  • Response Calculation: The detector calculates a "sum response" (SR) and a "difference response" (DR). The SR is designed to be high at chess-board vertices by summing intensities of opposing samples and subtracting intensities of samples at 90-degree intervals. The DR is designed to be high at edges, by summing the absolute differences of opposing samples.
  • Noise and Edge Rejection: To further refine the detection and reject false positives, a "mean response" is calculated by comparing the local intensity within the sampling ring to the intensity of the surrounding pixels. This helps distinguish true vertices from uniform stripes. The final response is a combination of SR, DR, and the mean response, designed to be high for true vertices and low for other features.
  • Orientation Labelling: An additional step assigns one of eight orientation labels to detected features, based on the rotation that maximizes the sum response. This is useful for subsequent processing and decoding.

Datasets

The paper evaluates its method using both synthetic and real-world data.

  • Synthetic Data: Images are generated by creating rotated chess-board patterns with varying levels of Gaussian noise and blur. This allows for controlled testing of the detector's robustness.
  • Real Data:
    • Flat Plate: A flat plate with a projected chess-board pattern is moved towards and away from cameras. This allows for evaluating the accuracy of reconstructed surfaces and camera calibration.
  • Cylinder: A cylinder with a projected chess-board pattern is used for a more challenging reconstruction scenario, testing the detector's performance on a non-planar surface with significant perspective distortion.

Key Insights and Ideas

  • Beyond Simple Thresholding: The paper argues against simple binary thresholding for feature detection, advocating for a continuous strength measure that allows for more flexible post-processing and decision-making.
  • Rotationally Invariant Sampling: The circular sampling strategy is key to achieving robustness against feature rotation, a common challenge in computer vision.
  • Combined Response for Robustness: The combination of sum, difference, and mean responses effectively discriminates between true vertices, edges, and uniform regions, significantly reducing false positives.
  • Pre-blurring Benefits: While the ChESS detector is robust, applying a slight pre-blur to the input image can further enhance its resilience to noise, especially in low-light conditions.
  • Efficiency is Key for Real-time: The detector's design prioritizes computational efficiency, making it suitable for real-time video processing.

Key Results

  • Accuracy and Robustness: ChESS consistently outperforms or matches the performance of established detectors like Harris and Stephens and SUSAN in terms of accuracy and robustness to noise and rotation, especially in challenging conditions.
  • Speed: The ChESS detector is significantly faster than the Harris detector, achieving approximately 40% speedup. Its SIMD-optimized version can process over 700 VGA frames per second.
  • Reconstruction Quality: Experiments with flat plate and cylinder reconstructions show that ChESS leads to more accurate and stable fits compared to other detectors, particularly in noisy or distorted scenarios.
  • Low False Positive Rate: The method demonstrates a low rate of false positives, especially when compared to general-purpose corner detectors.

Limitations and Future Work

  • Perspective Invariance: The authors acknowledge that the detector is not fully perspective-invariant. While it handles distortions, extreme perspective can still affect the response strength.
  • No Intensity/Contrast Normalization: The detector does not inherently normalize for intensity or contrast variations across the image, although a "neighbourhood comparison" step helps mitigate this.
  • Future Work: The paper suggests extending the ChESS approach to detect higher-order intersection features (beyond simple corners) and exploring its use with coded grids for more complex applications. The authors also note the potential for further optimization using SIMD instructions.

Summary updated

BibTeX
@article{bennett2014chess,
  author = {Stuart Bennett and Joan Lasenby},
  title = {{ChESS - Quick and robust detection of chess-board features}},
  journal = {CVIU},
  year = {2014},
  doi = {10.1016/j.cviu.2013.10.008},
}
Figure