PaperlayerView exact revision on arXiv

Source-linked AI summary

BlobBoards: Robust Markers for Accurate Pose

James Pritts, Till Sittart, Hendrik Sauer, Silja Janßen, Felix Seegräber, David Nakath, Kevin Köser

arXiv:2608.28830v1cs.CV

TL;DR

Existing fiducial systems limit feature density and can fail under partial visibility, motivating a marker that remains identifiable and pose-constraining across scale and viewpoint changes. BlobBoards use dense, multi-scale Gaussian blobs with learned descriptors and spatial verification for joint detection, identification, and pose estimation. They outperform the evaluated tag systems in detection, translation accuracy, and occlusion robustness, while reducing large-rotation failures.

  • Problem

    Conventional fiducials limit keypoint density and can become ambiguous or fail under partial visibility, motivating robust references for pose estimation and identification.

  • Method

    BlobBoards densely distribute multi-scale Gaussian blobs and use learned local descriptors with spatial verification to identify boards and estimate pose jointly.

  • Results

    BlobBoards outperform AprilTag and ArUco across detection rate, translation accuracy, large-rotation failures, and occlusion robustness in motion-capture evaluations.

  • Takeaways & Limitations

    Dense distributed features let BlobBoards retain accurate pose estimates as visibility decreases, including 69% detection at 50% occlusion.

  • Takeaways & Limitations

    Severe foreshortening can make descriptor matching unreliable where real training tracks are sparse, despite strong pose accuracy for successfully identified boards.

Abstract

from arXiv · show

We propose BlobBoards, a fiducial marker system comprising a dense, multi-scale field of Gaussian blobs and a feature-based pipeline for joint detection, identification, and pose estimation. Each board is registered from hundreds of blob features whose dense spatial coverage constrains pose, while multiple scales preserve detectability across large changes in focal length, distance, and obliquity. Learned local descriptors are matched to the reference pattern and spatially verified, so the correspondences determine pose and certify identity. Against motion-capture ground truth, BlobBoards achieve median translation errors of 3.6-5.0 mm, reducing AprilTag's median translation error by 89% on small boards and 70% on large ones. They also produce far fewer large-rotation failures than state-of-the-art tag systems. BlobBoards achieve the highest detection rate, 80% versus 74% for AprilTag and 58% for ArUco, with the largest margin on the smallest markers. Under 50% occlusion, they still detect 69% of boards with essentially unchanged median translation error, while AprilTag and ArUco detect none. In experiments BlobBoards give state-of-the-art detection rate, pose accuracy and occlusion robustness.

Introduction

BlobBoards address the limited keypoint density and partial-visibility weaknesses of conventional fiducials with dense, multi-scale Gaussian-blob patterns and feature-based registration. Their design supports accurate pose estimation, identification, and graceful degradation under occlusion.

  • Motivation: Conventional coded fiducials provide only four corner measurements per marker, so measurement density falls as marker size increases.Occluding the segmented boundary can disable the marker outright.
  • Design: BlobBoards use 213–1027 densely distributed Gaussian blobs across 4–12 cm boards, making measurement density a property of the pattern rather than board size.Multiple blob scales preserve feasible detection across changes in focal length, distance, and viewpoint.
  • Pipeline: Affine-adapted blobs are described by surrounding texture, then descriptor matches are spatially verified through RANSAC registration.The same correspondences simultaneously identify and register the board.
  • Results: Against motion-capture ground truth, BlobBoards improve pose accuracy and detection over AprilTag and ArUco, with larger advantages for smaller and more oblique markers.Under occlusion, visible-area loss removes local features progressively rather than disabling the marker.
  • Design: Gaussian blobs provide scale-covariant localization with scale-independent peak response, while compact support enables dense packing of accurately localized features.The maximum response is A/2 independent of blob scale σ0.
  • Generation: BlobBoard generation greedily packs blobs from multiple scales with disjoint supports, balancing density, scale coverage, and printability through configurable parameters.The generator outputs both the printable image and reference geometry for registration.

3 The BlobBoard Pipeline

BlobBoards use dense, multi-scale blob features and a feature-matching pipeline to detect, identify, and estimate board pose. The pipeline adapts affine blob frames, canonicalizes local texture, describes and matches features, and spatially verifies correspondences geometrically.

  • Detect & Adapt: BlobBoards detect candidates in ASIFT scale space, adapt them to anisotropic ellipses, and return refined elliptic frames for canonicalization.Affine warps provide seed shape estimates; IRLS refines each valid detection on the GPU.
  • Canonicalize & Describe: Whitening and log-polar resampling remove local affine deformation and encode residual rotation and scale as translations in a fixed 64×64 patch.The surrounding board texture, rather than the blob’s identical radial profile, provides discriminative information.
  • Canonicalize & Describe: A learned HardNet-style network maps each canonical patch to a 128-D unit-norm descriptor, with angular max pooling for rotation handling.The network uses 5×5 convolutions and a dense layer to produce the descriptor.
  • Match & Verify: Descriptor matching uses precomputed reference galleries, while spatial verification registers candidate boards and supports matching at O(NBMD) similarity cost.Reference descriptors are computed once, and verification fits board hypotheses from image correspondences.
  • Match & Verify: Correspondences are claimed in rounds so accepted boards remove their detections before remaining candidates are rematched one-to-one.This prevents false hypotheses from borrowing support explained by an already accepted board.

4 Evaluation Protocol

The evaluation compares BlobBoards with AprilTag and ArUco on matched marker panels observed across distances, viewpoints, and scales. Motion capture supplies pose ground truth, while recall curves expose detection-limited performance and calibration limits constrain rotation comparisons.

  • Experimental Setup: BlobBoards, AprilTag 3, and ArUco are evaluated on matched rigid panels carrying 30 markers at small, medium, and large physical scales.Pattern area is matched across methods, while BlobBoards provide hundreds of blobs versus four tag corners.
  • Experimental Setup: Captures span 1, 2, and 3 m distances and nominal panel rotations from −90° to +90° about fronto-parallel, with actual obliquity obtained from motion capture.The three marker types are captured at matched distances and viewpoints.
  • Ground-Truth Calibration: Motion-capture poses are converted into the camera frame using shared camera-rig and board-to-panel transforms estimated through robot–world/hand–eye calibration.Board-specific and shared transforms are recovered jointly from relative motions and linear least squares.
  • Evaluation Metrics: Figure 6 reports empirical recall CDFs for translation and rotation error by marker size, with missed detections limiting each curve’s saturation level.Columns represent 4 cm, 6 cm, and 12 cm markers; rows represent translation and rotation error.
  • Ground-Truth Calibration: Calibration residuals are 0.68 mm for translation and 0.46° for rotation, so rotation comparisons are restricted to large flip errors.Median rotation differences are smaller than the ground-truth resolution, whereas translation error remains well above calibration uncertainty.

5 Results

BlobBoards improve pose accuracy and detection relative to tag baselines, while remaining substantially more robust under occlusion. Their remaining failures concentrate in severe foreshortening and border-crossing occlusion settings.

  • Pose accuracy: 3.6, 3.7, and 5.0 mm are BlobBoards’ median translation errors for small, medium, and large boards.Relative to AprilTag, BlobBoards reduce translation error by 9.1×, 6.9×, and 3.3× across those sizes.
  • Detection: 80% overall probability of detection exceeds AprilTag’s 74% and ArUco’s 58%.For the smallest markers, detection rates are 78%, 69%, and 45%, respectively.
  • Failure modes: Most remaining BlobBoard failures under severe foreshortening occur during descriptor matching, despite successful blob detection and affine adaptation.The paper attributes this to sparse real training tracks; successfully identified boards retain strong pose accuracy.
  • Evaluation: The evaluation masks a controlled fraction of each marker in reference coordinates and warps it with ground-truth homography for matched occlusion across methods.This design supports direct comparison of BlobBoards, AprilTag, and ArUco under the same occluded area.
  • Occlusion robustness: At 10% occlusion, BlobBoards retain 0.8 probability of detection, whereas AprilTag falls to 0.11 and ArUco to 0.23.From 20% occlusion onward AprilTag reports no detections, and ArUco reaches zero by 30%.
  • Occlusion robustness: At 50% occlusion, BlobBoards detect 69% of markers with median errors of 1.05° and 3.9 mm.Some poses remain recoverable even at 90% occlusion, while translation error stays between 3.6 and 4.0 mm across the sweep.
  • Rotation accuracy: 33 planar-pose flips remain for BlobBoards after likelihood-ratio correction, versus 124 for AprilTag and 152 for ArUco.The correction reduces BlobBoard flips from 45 to 33 without changing detections, correspondences, or calibration.

6 Implementation

BlobBoards use GPU-batched Julia kernels, whereas the baselines run on CPUs; the implementation comparison therefore reports measured timings rather than hardware-normalized speed.

  • Runtime: 2.03 ± 2.18 s per image is BlobBoards’ measured warm runtime on 12-megapixel images containing 30 boards.The measurement uses an NVIDIA RTX 4070 and is reported against CPU baseline timings without hardware normalization.
  • Runtime: 0.11 ± 0.01 s and 0.04 ± 0.01 s per image are the corresponding AprilTag and ArUco CPU runtimes.AprilTag runs through dt-apriltags, while ArUco uses cv.aruco with default detector parameters.
  • Pose disambiguation: 33 planar-pose flips remain after the likelihood-ratio test, down from 45 with identical detections, correspondences, and calibration.The tag comparison uses IPPE’s lower-reprojection-error pose from four corners, while BlobBoards use hundreds of verified correspondences.

7 Conclusion

BlobBoards combine dense, scale-distributed Gaussian features with feature matching, spatial verification, and appearance certification to estimate pose and identify boards. Against tag systems, they improve detection and translation accuracy while reducing large-rotation failures, especially for difficult viewing conditions.

  • BlobBoards use feature matching, spatial verification, appearance certification, affine adaptation, and scale-marginalized likelihood testing to resolve pose ambiguity.
  • BlobBoards outperform state-of-the-art tag systems in detection rate and translation accuracy while substantially reducing large-rotation failures.
  • The gains are largest for the smallest, most oblique, and most occluded boards, where conventional tags are most fragile.
  • The evaluation uses motion-capture ground truth, with the setup provided by the Kurt Semm Centre and funded through specified institutional programs.

Supplementary Material

The paper is followed by supplementary material whose sections, figures, tables, equations, and algorithms use the S prefix for citation.

  • Supplementary sections, figures, tables, equations, and algorithms are labeled with the S prefix.

S1 Board Generation

BlobBoards are generated by placing non-overlapping Gaussian blobs from largest to smallest, then rasterising the resulting set into a board pattern. Shared parameters across physical sizes make blob count grow with board area, unlike tags that contribute four corners.

  • Blobs are placed largest-first by rejection sampling with disjoint supports, because small-first placement leaves less room for large blobs.
  • The generation procedure samples blob centres within the inset board and accepts candidates whose support discs do not overlap existing blobs.
  • Accepted blobs are rasterised over their supports in one pass on a white background, enabled by disjoint supports.
  • All three physical sizes share one parameter set, while pattern extent changes and blob count therefore grows with area.
  • The evaluation uses 18 small, 8 medium, and 4 large boards, with blob-count ranges of 202–226, 365–390, and 1001–1032.

S3 Finite-Support Scale Correction

Finite Gaussian support perturbs scale-space response, but a correction based on the support-dependent bias restores the reference scale under the continuous model.

  • Truncating the Gaussian at c and removing pedestal p perturbs the scale-space response, with ρ, κ, and q parameterizing the correction.
  • The response extremum occurs at β(c)σ0, where β(c) captures the support-dependent pedestal scale bias.
  • Because β depends only on c, the truncated response remains scale-covariant with peak magnitude independent of σ0.
  • Dividing each detected scale by β(c) recovers σ0 exactly under the continuous model.
  • As c approaches infinity, the response recovers the ideal Laplacian-of-Gaussian response.

S6 Pipeline

The pipeline processes candidate boards through shared image stages, then performs per-board matching, spatial verification, pose disambiguation, certification, and detection claiming. It evaluates comparable marker panels across scales, distances, and obliquities using motion-capture-tracked poses.

  • Stages (1)–(4) run once per image, while stages (5)–(8) run per board and compete for the same detections.
  • The algorithm matches gallery descriptors, resolves one-to-one correspondences, fits pose with LO-RANSAC, resolves mirror ambiguity, and certifies candidates using inlier support and appearance.
  • The best-supported candidate is committed, and detections inside its projected board region are claimed before subsequent boards are processed.
  • Experiments compare BlobBoards, AprilTag, and ArUco on matched rigid panels containing 30 markers at three physical scales.

S8 Hand–Eye Calibration Diagnostics

The calibration diagnostics describe independent, overdetermined hand–eye fits and distinguish absorbable constant system bias from performance degradation that varies with scale, distance, or obliquity. Supplementary figures characterize recall, unoccluded performance, and the deterministic occlusion construction used elsewhere in the evaluation.

  • 41, 22 and 15 usable calibration frames were used for BlobBoards, AprilTag and ArUco, versus 56, 57 and 57 evaluation frames, respectively.
  • Recall curves are pooled over marker sizes, with missed detections excluded from the numerator, so each curve saturates at its detection probability.
  • Up to 30 boards across 15–41 calibration frames constrain 6+6×30 parameters in a heavily overdetermined fit spanning evaluation distances and obliquities.
  • Only detector errors constant across the calibration span can be absorbed as rigid-mount error; scale-, distance- and obliquity-dependent degradation remains in reported results.
  • Unoccluded performance is binned by obliquity and camera distance, reporting per-bin medians and interquartile ranges while omitting bins with fewer than five detections.
  • The occlusion sweep uses one deterministic, pose-warped region per marker with exactly matched covered fraction at every level.

S13 Occlusion by Floor-Plan Level

Floor-plan views compare detection coverage and pose accuracy at 10%, 50%, and 70% occlusion using station-wise detection fractions and median translation and rotation errors. BlobBoards retain coverage as tag baselines become gray, with the separation already visible at 10% and complete tag-plan failure at 70%.

  • 10% occlusion: 10% occlusion reveals the design split: border-crossing occluders break the single quad used by AprilTag and ArUco, while BlobBoards lose only covered blobs.
  • 50% occlusion: At 50% occlusion, neither tag baseline returns poses anywhere on the rig stations, while the BlobBoards plan remains intact.
  • 10% occlusion: At 10% occlusion, AprilTag and ArUco plans are already largely gray, whereas BlobBoards retain coverage across the capture floor.
  • Each floor plan encodes station-wise detection fraction, median translation error, and median rotation error, with gray wedges marking zero detections.
  • 70% occlusion: At 70% occlusion, both tag plans are uniformly gray, while BlobBoards retain coverage with per-station errors in their unoccluded range.
Loading 2608.28830v1…