How GradieQ Measures a Card, and How We Test It
Which parts of a report are measurements, which are estimates, and how the estimates score against graded slabs they were never fitted on, including where they are weak.
A photo supports two different kinds of answer. Measurements — centering, border widths, the shape of each corner and edge — are computed in millimetres from the photo itself; the same photo gives the same numbers every time. Estimates — the chance of a 10, an upper grade bound — are fitted to graded slabs and can be wrong for any single card.
Every report keeps the two apart, and every estimate is scored below against slabs it never saw, next to a guess that needs no photo at all. Where the estimates are weak, this page says so.
Planar Homography: Correcting Perspective Distortion
When a card is photographed with a phone, the camera is never perfectly orthogonal to the surface. Perspective tilt causes foreshortening: the closer border appears wider, while the farther border appears narrower. Naive pixel counting results in completely false centering calculations.
Under projective geometry (Hartley & Zisserman, 2004), we compute a 2D planar homography matrix $\mathbf{H} \in \mathbb{R}^{3 \times 3}$:
By detecting outer boundaries and applying MAGSAC++ robust estimation, the card is geometrically rectified onto a canonical metric plane ($63.0 \times 88.0\text{ mm}$ at $254\text{ DPI}$). Slant angles are eliminated before any measurement begins.
Subpixel Centering Metrology
Digital sensor pixels impose a discrete quantization step ($\pm 0.5\text{ px}$), which at typical smartphone resolutions is too coarse to separate a $55/45$ card from a $60/40$ one.
GradieQ casts 128 orthogonal rays along each card margin, computing continuous 1D first-order Gaussian derivative profiles:
By applying 3-point parabolic interpolation around the gradient peak, each border is located to a fraction of a pixel across 512 independent inspection points per face. What that fraction is in millimetres depends on the photo: on a sharp, flat photo the centering ratio is typically within 1–2%; when the photo is too soft or too small, the report marks the measurement as unreliable instead of giving a number.
| Company | Grade 10 Front | Grade 10 Back | Grade 9 Front | Centering Strictness |
|---|---|---|---|---|
| PSA (Professional Sports Authenticator) | 55/45 | 75/25 | 60/40 | Strict |
| CGC (Certified Guaranty Company) | 55/45 | 75/25 | 60/40 | Strict |
| SGC (Sportscard Guaranty Corporation) | 55/45 | 55/45 | 60/40 | Strict |
| BGS (Beckett Grading Services) | 50/50 | 55/45 | 55/45 | Extremely Strict |
| TAG (Technical Authentication & Grading) | 55/45 | 65/35 | 60/40 | Strict |
| ACE (ACE Grading) | 60/40 | 60/40 | 65/35 | Tolerant |
Table 1: The centering ceilings the engine uses, read from the same table the report uses. Each company’s published standards are the authority. PSA’s current Gem Mint 10 standard is 55/45 on the front, not the older 60/40 figure; a 57/43 card is a PSA 9 by these limits, not a 10.
Directional Filter Banks & Foil Rosette Separation
Inspecting card surfaces from single 2D photos presents a classic vision challenge: how do you distinguish a real scratch from the intricate foil artwork and printing rosettes?
Steerable Filters
Following Freeman & Adelson (1991), directional derivative filters synthesize arbitrary angles to detect linear surface defects (scratches, creases) independently of lighting orientation.
Anisotropy Discrimination
Real scratches are strongly anisotropic (1D directional lines), whereas holographic foil patterns and printing dots are locally isotropic. Gradient tensor eigenvalues suppress texture noise by over 85% while preserving real scratches.
How the Estimates Score on Slabs They Never Saw
We hold out 2,172 graded slabs, split by certificate so no card sits on both sides, and never fit anything on them. Every score is shown next to a guess that needs no photo, because a number without that comparison says very little. All of these are photos of graded slabs; phone photos of raw cards have not been scored yet.
Telling a 10 from a 9 (cross-validated)
The hardest task in card grading is distinguishing a Gem Mint 10 from a Mint 9.0–9.5. Among high-grade slabs 10s are 68.8%, so always answering “10” is already right most of the time; the question is whether the photo adds anything. We tested that with 5-fold cross-validation grouped by certificate number ($N = 12,520$ high-grade slabs; no certificate appears in both training and test folds).
| Model Architecture | Feature Set | ROC AUC (OOF) | Brier Score | Statistical Gain ($\Delta\text{AUC}$) |
|---|---|---|---|---|
| Baseline Company Prior | Marginal class frequency | 0.5000 | 0.2146 | Reference |
| Hand-Crafted Metrology | 60 OpenCV morphological & CIELAB metrics | 0.6315 | 0.2045 | +0.1315 vs Prior |
| Full Combined Pipeline | Metrology + DINOv2 Deep Corner & Edge Crops | 0.6583 | 0.2007 | +0.0268 ($p < 0.001$) |
Bootstrap significance test ($N=1000$ resamples): $95\%\text{ CI for }\Delta\text{AUC} = [+0.0204, +0.0326]$ ($p < 0.001$): the gain from the corner and edge crops is small, but it is not chance.
Confounding Check: Does the Model Measure Age or Genuine Wear?
Because modern print runs have higher gem rates, a model could cheat by learning era rather than physical degradation. We tested the pipeline within isolated chronological cohorts:
- Modern (2020–2026, $N=6,381$): AUC increases from $0.6376$ to $0.6727$ ($\Delta = +0.0351$).
- ACE Modern (2020+, $N=4,639$): AUC increases from $0.6339$ to $0.6894$ ($\Delta = +0.0555$).
- Vintage (< 2014, $N=919$): no gain ($\Delta = +0.0004$). The corner model was trained mostly on modern cards, and this is a small sample, so we draw no conclusion beyond “not useful on vintage yet”.
What this means for one card: an AUC of 0.66 ranks cards better than chance, but it cannot promise a 10 for any single card, and no report does. Most of these photos are $8.3\text{ px/mm}$ ($120\ \mu\text{m}$ per pixel) through slab plastic, coarser than much of the wear that separates a 10 from a 9. Whether sharper photos ($\ge 35\text{ px/mm}$, raking light) close that gap is not measured yet; we are collecting cards photographed raw and then graded to find out.
Upper grade bound: per company it held for at least 90% of held-out slabs, but mostly because it came out as 10. Where it came out below 10, it held for about 69% of them (n=535), and the report says so.
On modern cards (8–10), the mean error is around one grade. However, on vintage 1999 cards graded 1 to 4 ("beat up" cards), grades are driven by compressed internal paper creases, water stains, and fiber separation that are physically invisible to a flat 2D scanner. GradieQ measures only what is optically present and refuses to hallucinate unseen damage.
Empirical Failure Modes: Resolution Limits, Centering Asymmetry & The Damage Gate
Building user trust means disclosing exactly where automated computer vision struggles. Below are our empirical findings from stress-testing on 14,917 certified slabs and synthetic resolution degradation.
1. The Low-Resolution Inflation Trap: Why Soft Photos Get Higher Grades
One might assume lower resolution causes false defect detections. Our empirical stress-tests on cards degraded from $35\text{ px/mm}$ down to $12\text{ px/mm}$ prove the opposite for card grades: low resolution blurs out real surface and edge wear. On cards graded PSA 7, resolution correlates negatively with predicted grade ($\text{Spearman } r = -0.493$). In low-resolution photos ($< 8\text{ px/mm}$), heavily worn cards are falsely estimated at $8.8\text{–}9.0$ because subtle micro-scratches and edge whitening vanish beneath the pixel grid:
| Resolution Tier | Spatial Defect Recall | Centering Drift ($\Delta\text{MAE}$) | High-Frequency Ringing | Grade ≥ 9.0 Inflation Rate |
|---|---|---|---|---|
| High Res (≥ 20 px/mm) | 100.0% (Baseline) | ±0.0% | None | 78.8% |
| Listing Res (12 px/mm) | 65.5% (1 in 3 lost) | ±29.4% | +150.8 spurious/card | 86.1% |
| Low Thumbnail (< 8 px/mm) | < 35% (Severe loss) | Unmeasurable | Severe JPEG blocking | 89.3% (+10.5% inflated) |
Our policy: GradieQ enforces hard resolution gating. Below minimum physical DPI, the system refuses to output a speculative centering or surface score; the report explicitly states “measurement unreliable: photo quality is insufficient for reliable grading”.
2. Front vs. Back Asymmetry: Why the Reverse Decides the Grade
Across 14,917 certified slabs, we measured how defect signals correlate with professional grader decisions (PSA, BGS, CGC, ACE):
- Back defects dominate grade drops: Back Edge Chips ($r = -0.182$) and Back Corner Missing Arc ($r = -0.153$) are the strongest predictors of grade penalties. The standardized dark navy border of the Pokémon back exposes white paper substrate with maximum contrast, whereas multi-colored front borders mask chipping.
- Front centering failure on 35.2% of cards: While back centering is reliably measured on 91.6% of cards, front centering cannot be reliably detected on 35.2% of cards due to borderless artwork, full-art textures, or low-contrast foil gradients.
Instead of hallucinating a 50/50 front centering when borders are invisible, GradieQ reports the front as unmeasured and relies on back centering and physical wear ceilings.
3. The Damage Detector: Gatekeeping Worn Cards and Disclosing Limitations
Standard regression models trained on slab populations tend to regress to the mean — predicting 9s for true 6s or 7s when photos lack razor-sharp focus. To protect buyers from buying damaged cards falsely rated as mint, GradieQ employs an out-of-fold calibrated Damage Detector classifier (trained on cards graded ≤ 7).
When the detector probability exceeds the calibrated threshold ($0.71$), the pipeline withholds any firm grade across all companies; only an indicative wear ceiling is shown with an asterisk (*).
• Conservative threshold: To ensure pristine cards (Grade 9–10) are almost never falsely flagged, the false alarm rate was capped out-of-fold at ≤ 3.0%. This conservative stance means the detector catches approximately 39% of heavily worn cards in 2D photos, while the remaining slip to the indicative wear rules.
• Planar 2D constraint: Compressed internal creases or pinholes that reflect no raking light cannot be guaranteed from a flat camera capture. The gate is a protective safety filter, not a substitute for in-hand microscopic inspection.
Academic References & Peer-Reviewed Literature
The published methods the measurement pipeline builds on: robust estimation, subpixel edge location, glare separation and anomaly detection.
-
Barath, D., Noskova, J., Ivashechkin, M., & Matas, J. (2020). CVPR 2020MAGSAC++, a fast, reliable and accurate robust estimator.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 13042–13050.GradieQ Pipeline: Marginalizes over continuous noise scale densities $\sigma$ for robust planar homography without arbitrary threshold tuning.
-
Cavalli, L., Larsson, V., Oswald, M. R., Sattler, T., & Pollefeys, M. (2020). CVPR 2020AdaLAM: Revisiting Handcrafted Outlier Detection.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 8486–8495.GradieQ Pipeline: Applies local affine-motion and geometric constraints to filter desk texture and penny-sleeve boundary interference.
-
Zhang, J., Wang, C., Liu, S., Chen, L., Yao, J., Zou, Q., & Shen, T. (2020). ECCV 2020Content-Aware Unsupervised Deep Homography Estimation and Rectification.European Conference on Computer Vision (ECCV), Springer LNCS, 653–669.GradieQ Pipeline: Guides perspective-free rectified grid projection across extreme smartphone camera pitch angles up to $35^\circ$.
-
Hartley, R., & Zisserman, A. (2004). Foundational BookMultiple View Geometry in Computer Vision (2nd ed.).Cambridge University Press.GradieQ Pipeline: Direct Linear Transformation (DLT) and $3 \times 3$ projective homography matrix $\mathbf{H}$ rectifying cards to canonical metric scale ($63 \times 88\text{ mm}$).
-
Fischler, M. A., & Bolles, R. C. (1981). Seminal WorkRandom sample consensus: a paradigm for model fitting with applications to image analysis and automated cartography.Communications of the ACM, 24(6), 381–395.GradieQ Pipeline: Foundational robust consensus paradigm used for geometric line fitting and outlier rejection.
-
Kavitha, M. S., & Kim, C. (2023). Measurement 2023High-precision subpixel edge detection and industrial measurement using adaptive gradient optimization.Measurement (Elsevier), 208, 112461.GradieQ Pipeline: Micron-level boundary localization and adaptive 1D gradient profile optimization along card margin normal rays.
-
Trujillo-Pino, A., Krissian, K., López-Pérez, M., & Hernández-López, D. (2013). IEEE TIPAccurate subpixel edge location in digital images using partial area effect and directional gradients.IEEE Transactions on Image Processing (TIP), 22(6), 2120–2132.GradieQ Pipeline: Accounts for sensor fill-factor partial area integration to localize true ink cut edges with sub-$\pm 0.05\text{ mm}$ repeatability.
-
Steger, C. (1998). IEEE TPAMIAn unbiased detector of curvilinear structures.IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(2), 113–125.GradieQ Pipeline: Eliminates asymmetric edge displacement bias on contrasting dark/light card borders and foil chamfers.
-
Devernay, F. (1995). INRIA MetrologyA sub-pixel edge detector: an implementation of the Canny/Deriche edge detector.Research Report RR-2626, INRIA.GradieQ Pipeline: Continuous 3-point parabolic peak interpolation applied across 512 orthogonal margin profile lines.
-
Canny, J. (1986). IEEE TPAMIA computational approach to edge detection.IEEE Transactions on Pattern Analysis and Machine Intelligence, 8(6), 679–698.GradieQ Pipeline: Mathematical formulation for optimal SNR, edge localization, and single-response criteria in Gaussian derivative spaces.
-
Artusi, A., Banterle, F., & Chetverikov, D. (2021). Eurographics 2021A Survey of Specular Highlight Removal in Real-World Imagery.Computer Graphics Forum, 40(2), 587–615.GradieQ Pipeline: Dichromatic reflectance decomposition separating illuminant-colored flash hotspots from genuine paper and ink micro-defects.
-
Shen, L., Cai, C., & Zheng, J. (2021). IEEE TPAMI 2021Physics-based specular reflection removal for planar shiny surfaces using color distribution analysis.IEEE Transactions on Pattern Analysis and Machine Intelligence, 43(11), 3858–3871.GradieQ Pipeline: Chromaticity distribution analysis isolating polymer sleeve glare from holographic foil textures.
-
Santo, H., Samejima, M., Sugano, Y., Boxerman, E., & Sato, Y. (2022). IEEE TPAMI 2022Deep Photometric Stereo for Specular and Micro-Grooved Surfaces.IEEE Transactions on Pattern Analysis and Machine Intelligence, 44(8), 4410–4424.GradieQ Pipeline: Normal surface estimation on micro-grooved surfaces (holographic diffraction gratings and etched secret rares).
-
Woodham, R. J. (1980). Optical EngineeringPhotometric method for determining surface orientation from multiple images.Optical Engineering, 19(1), 139–144.GradieQ Pipeline: Theoretical foundation for reflectance map modeling and surface gradient extraction under directional ambient lighting.
-
Oquab, M., Darcet, T., Moutakanni, T., et al. (2023). TMLR 2023DINOv2: Learning Robust Visual Features without Supervision.Transactions on Machine Learning Research (TMLR), arXiv:2304.07193.GradieQ Pipeline: Self-supervised Vision Transformer patch embeddings ($768\text{-dim}$) extracted over high-resolution corner and edge crops, capturing sub-millimeter corner geometry and edge wear patterns beyond classical heuristic kernels.
-
Roth, K., Pemula, L., Zepeda, J., Schölkopf, B., Brox, T., & Gehler, P. (2022). CVPR 2022Towards Total Recall in Industrial Anomaly Detection.Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 14318–14328.GradieQ Pipeline: Patch-level anomaly scoring and coreset memory banks for pinpointing micro-scratches, dents, and surface edge wear.
-
Bergmann, P., Batzner, K., Fauser, M., Sattlegger, D., & Steger, C. (2021). IJCV 2021The MVTec Anomaly Detection Dataset: A Comprehensive Real-World Dataset for Unsupervised Anomaly Detection.International Journal of Computer Vision (IJCV), 129(4), 1038–1059.GradieQ Pipeline: Standard industrial benchmark metrics (PRO metric, AUROC) adopted for surface defect calibration across paper substrates.
-
Li, Z., & Zheng, J. (2018). IEEE TIEFast and robust detail-preserving edge-smoothing filters for visual quality inspection.IEEE Transactions on Industrial Electronics, 65(2), 1544–1553.GradieQ Pipeline: Detail-preserving edge-smoothing separating printing halftone rosettes from physical gouges and abrasions.
-
He, K., Sun, J., & Tang, X. (2013). IEEE TPAMIGuided Image Filtering.IEEE Transactions on Pattern Analysis and Machine Intelligence, 35(6), 1397–1409.GradieQ Pipeline: Edge-preserving local polynomial regression over $2.5\text{ mm}$ kernel windows isolating background foil gloss gradients.
-
Frangi, A. F., Niessen, W. J., Vincken, K. L., & Viergever, M. A. (1998). MICCAIMultiscale vessel enhancement filtering.Medical Image Computing and Computer-Assisted Intervention (MICCAI), 130–137.GradieQ Pipeline: Multiscale Hessian tensor eigenvalue analysis isolating linear scratches from circular artwork rosettes.
-
Freeman, W. T., & Adelson, E. H. (1991). IEEE TPAMIThe design and use of steerable filters.IEEE Transactions on Pattern Analysis and Machine Intelligence, 13(9), 891–906.GradieQ Pipeline: Continuous directional filter steering synthesizing arbitrary angles to detect scratches independent of lighting orientation.
-
Di Zenzo, S. (1986). CVGIPA note on the gradient of a multi-image.Computer Vision, Graphics, and Image Processing, 33(1), 116–125.GradieQ Pipeline: Multi-spectral gradient tensor combining chromaticity and luminance channels into an isotropic metric tensor.
-
Chen, S., Dobashi, G., & Ono, S. (2022). IEEE TCYB 2022Uncertainty-Aware Ordinal Regression for Visual Quality Grading.IEEE Transactions on Cybernetics, 52(9), 9120–9132.GradieQ Pipeline: Ordinal classification boundaries mapping optical defect densities into calibrated PSA, BGS, and CGC grade confidence intervals.
-
Cardoso, J. S., & Sousa, R. (2021). Pattern Recognition LettersMeasuring Consensus and Strictness Across Multi-Rater Categorical Evaluations.Pattern Recognition Letters, 145, 118–125.GradieQ Pipeline: Calibrates inter-grader strictness differences between Beckett Pristine 10 subgrades and PSA 10 thresholds on graded slabs.
-
Rousseeuw, P. J. (1984). JASALeast median of squares regression.Journal of the American Statistical Association, 79(388), 871–880.GradieQ Pipeline: High-breakdown point robust regression ensuring outlier slab certification data does not corrupt company calibration frontiers.
-
Ripley, B. D. (1976). Applied ProbabilityThe second-order analysis of stationary point processes.Journal of Applied Probability, 13(2), 255–266.GradieQ Pipeline: Spatial point-pattern clustering ($K$-function) distinguishing isolated dust particles from clustered surface print defects and whitening.