Definition

A chessboard X-corner is the saddle point at the meeting of four square cells in a planar checkerboard calibration target — locally, an ideal X-shaped intensity pattern with two bright quadrants opposite two dark quadrants. Detecting these inner corners and assembling them into a labelled grid is the foundational input to every classical camera-calibration pipeline (Zhang's planar method, Geiger's single-shot calibration, ROS / Kalibr / OpenCV camera calibrators).

Although the geometry of the target is fixed and well known, robust detection is non-trivial: practical images suffer from defocus and motion blur, severe perspective and lens distortion, partial occlusion, glare, low contrast in multispectral or thermal-IR captures, and background clutter that can produce convincing X-shaped artefacts. Twenty-five years of method development have converged on a small set of design axes; this survey organises the canonical methods along those axes and provides a decision table for picking one.

The four design axes

Every X-corner detector can be decomposed into four stages, and the choice at each stage is what differentiates one method from another:

1. Per-pixel corner response. A score image $R: \Omega \to \mathbb{R}$ where local maxima are X-corner candidates. Choices include the Hessian saddle response ($S = I_{xy}^2 - I_{xx}I_{yy}$), the four-quadrant convolution likelihood (Geiger), 16-pixel ring sampling (ChESS, Bennett 2013; FAST-derived), Bresenham-ring sign-alternation counting (Laureano), Hessian-saddle + structure-tensor (Shu, PuzzleBoard), gradient-magnitude centreline graph (ROCHADE), or a learned feature map — plain CNN (MATE, CCDN) or UNet 2D-Gaussian heatmap (CCS). For surveys see Hessian-saddle response.
2. Multi-scale strategy. None (single scale; ChESS, MATE), three fixed window sizes (Geiger), full image pyramid with per-corner level selection (Pyramidal blur-aware), or scale-space pyramid for blur invariance (Hillen GP enhancement on top of Geiger).
3. Structure recovery. The mechanism that promotes a set of candidate corners into a labelled grid. Choices: greedy energy-minimisation expansion from seed corners (Geiger), exact subgraph isomorphism against a model graph (OCPAD), topological grid recovery by quad/triangle filters and consistency rules (Shu, Laureano), de-Bruijn-coded position decoding (PuzzleBoard), or learned regression / classification post-processing (CCDN's NMS + k-means++).
4. Subpixel refinement. The mechanism that takes integer pixel candidates to floating-point coordinates. Hand-crafted choices: Hessian-Taylor solve (Chen 2005), cone-quadratic fit on the gradient-magnitude centreline (ROCHADE), gradient-orthogonality weighted least squares (Geiger), orientation-weighted DFT response (ChESS), Radon centroid (Duda), grayscale centroid (PuzzleBoard), nonlinear least-squares fit of an explicit blurred-corner model with seven parameters $(\mu, \upsilon, \alpha, \beta, \lambda, \kappa, \sigma)$ to the raw image patch and a built-in boxplot fit-quality self-check (Yang 2018). Learned choices: SVD-based Gaussian surface fit on the predicted heatmap, recovering both the sub-pixel mean $\mu$ and a confidence variance $\sigma$ (CCS). MATE and CCDN leave subpixel as a downstream concern. A unifying subpixel-corner-refinement concept page is a candidate next addition once a fourth mechanism converges.

Decision table

Picking one

The choice flows from operating regime, not from accuracy alone. The questions below are roughly ordered by importance.

Do you have multi-frame or video calibration with the camera held steady? Use Pyramidal blur-aware — its per-corner pyramid-level metadata is useful for autofocus and per-corner uncertainty, and it tops the F1 leaderboard on the abeles 2021 benchmark.

Are you running a single-shot multi-board rig (Kalibr-style, multiple targets in one image)? Use Geiger / libcbdetect. Its energy-minimisation structure recovery handles unknown numbers of boards in one pass and downstream tools (Kalibr, ROS) consume its output natively.

Do you need maximum sub-pixel accuracy and you know $(r \times c)$? Use ROCHADE for indoor / well-controlled imagery (0.0332 px reported), or Duda Radon corners under heavy blur where the centreline graph degrades.

Are your captures partially occluded or extending outside the frame? Two complementary approaches: OCPAD (combinatorial subgraph isomorphism — works without training) and GP enhancement (regression-based corner fill-in, works on any upstream detector's partial output). They are complementary; OCPAD recovers the partial-subgraph; GP fills in occluded or out-of-frame corners.

Are you working with multispectral, thermal-IR, or low-contrast endoscopic imagery? Pair Geiger with GP enhancement — the GP wrapper consistently improves Geiger's detection counts and corner accuracy on this class of imagery (Hillen 2023 §4).

Do you need a learned baseline? Use CCDN. It supersedes MATE on every reported metric. MATE remains relevant only as a historical baseline or for extreme parameter-budget situations (2,939 vs 16,301 weights).

Do you need an end-to-end learned calibration pipeline (detection + distortion correction + parameter estimation) with sub-pixel accuracy? Use CCS. The UNet heatmap detector with SVD-based Gaussian surface fitting delivers sub-pixel coordinates with confidence variance, the upstream CNN corrects radial distortion before detection, and the downstream image-level RANSAC stabilises Zhang-style intrinsic estimation (real-data RPE 0.37 px / STD 0.02 versus a Matlab Zhang baseline at 0.45 px / STD 0.10). The trade-off is a hard chessboard-only assumption and the need to match the target camera's distortion to the CCS training distribution.

Is fast hand-crafted scoring the priority and subpixel can be done downstream? Use ChESS. Single-pass per-pixel ring score, no graph stage; pair with a separate Hessian-Taylor or cone-quadratic refinement step.

Do you need sub-pixel refinement that operates on the raw image with built-in PnP outlier rejection? Use Yang sub-pixel corner fit. Seven-parameter blurred-corner LSQ over a $(2r+1)^2$ ROI without preprocessing; the per-corner RMSE drives a boxplot fit-quality self-check that down-weights unreliable corners before pose estimation. Designed as a refinement stage downstream of any pixel-level detector.

Do you need to identify each corner with an absolute integer coordinate without a manual hand-pick (multi-board, fragmented capture)? Use PuzzleBoard — its 501 × 501 de-Bruijn-coded pattern lets the detector decode each corner's absolute grid index from a $3 \times 3$ neighbourhood of binary codes.

Where the methods overlap

• Hessian saddle response. ChESS variants, ROCHADE, Shu, Laureano, PuzzleBoard, and indirectly the chen2005 line all build on $S = I_{xy}^2 - I_{xx}I_{yy}$ (with various sign conventions and smoothing choices). The unifying concept page is Hessian-saddle response.
• Topological grid recovery. Shu, Laureano, OCPAD, and PuzzleBoard all promote candidate corners to a grid by graph-topological constraints. The unifying concept page is topological grid recovery.
• Subpixel refinement. Several distinct mechanisms (Hessian-Taylor, cone-quadratic on a centreline graph, gradient-orthogonality WLS, orientation-weighted DFT, Radon centroid, grayscale centroid, parametric-model LSQ on raw patches) coexist; choice is largely independent of the detection stage and can be mixed across pipelines.

Open problems

• Robustness to extreme perspective. No current method handles fisheye-grade lens distortion combined with extreme tilt without an explicit pre-rectification step. PuzzleBoard's de-Bruijn decoding tolerates more distortion than grid-topology methods because it operates on local $3 \times 3$ binary codes, but at the cost of requiring its specific target.
• Multispectral / thermal IR. Hillen 2023 GP enhancement is the only published method evaluated explicitly on these modalities; the design space is otherwise dominated by visible-light evaluation.
• Occlusion + low contrast simultaneously. OCPAD requires a strong upstream graph; GP enhancement requires a partial board fragment. Neither bootstraps from heavy occlusion plus low contrast, where the upstream detector cannot return enough corners.
• Learning beyond CCDN. Modern transformer architectures and self-supervised pretraining (SuperPoint-style Homographic Adaptation) have not been widely applied to the chessboard X-corner problem. Open question: does the small specialised target make pretraining unnecessary, or would a larger model improve robustness on the multispectral / IR cases that GP enhancement currently fills?

Where this concept appears

Every chessboard X-corner detector page in the atlas lists this concept in related. The unifying sub-concepts — Hessian-saddle response and topological grid recovery — slice the design space differently and will be more useful for understanding why a particular method is structured the way it is.

References

1. S. Bennett, J. Lasenby. ChESS — Quick and Robust Detection of Chess-board Features. arXiv
   .5491, 2013.
2. A. Geiger, F. Moosmann, O. Car, B. Schuster. Automatic Camera and Range Sensor Calibration Using a Single Shot. IEEE ICRA, 2012.
3. S. Placht et al. ROCHADE: Robust Checkerboard Advanced Detection for Camera Calibration. ECCV, 2014.
4. F. Abeles. A Pyramidal Blur-Aware X-Corner Detector. IEEE ICCAR, 2021.
5. C. Shu, A. Brunton, M. A. Fiala. Topological Grid Recovery. (See atlas page for venue.)
6. G. T. Laureano, M. S. V. de Paiva, A. S. da Silva. A Topological-Approach for Detecting Chessboard Patterns. (See atlas page.)
7. P. Fürsattel et al. OCPAD — Occluded Checkerboard Pattern Detector. WACV, 2016.
8. J. M. Duda, K. Frese. Accurate Detection and Localisation of Checkerboard Corners for Calibration. BMVC, 2018.
9. S. Donné et al. MATE: Machine Learning for Adaptive Calibration Template Detection. MDPI Sensors, 2016.
10. B. Chen, C. Xiong, Q. Zhang. CCDN: Checkerboard Corner Detection Network for Robust Camera Calibration. arXiv
    .05097, 2023.
11. Y. Zhang, X. Zhao, D. Qian. Learning-Based Distortion Correction and Feature Detection for High Precision and Robust Camera Calibration. IEEE Robotics and Automation Letters 7(4)
    –10477, 2022.
12. P. Stelldinger, N. Schönherr, J. Biermann. PuzzleBoard. (See atlas page.)
13. M. Hillen et al. Enhanced Checkerboard Detection Using Gaussian Processes. MDPI Mathematics 11(22)
    , 2023.
14. T. Yang, Q. Zhao, X. Wang, Q. Zhou. Sub-Pixel Chessboard Corner Localization for Camera Calibration and Pose Estimation. MDPI Applied Sciences 8(11)
    , 2018.