cs.CVAug 28, 2026
SaveBlobBoards: Robust Markers for Accurate Pose
Organizations: Marine Data Science Kiel University
Abstract
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.
Figures & tables
Figure 1: All BlobBoards in the image are detected, identified, and registered against a -board gallery, so of its entries are absent confusers, and identification is still perfect. Colour distinguishes detections; the overlaid pattern shows registration accuracy across large variations in scale, distance, and obliquity.
Figure 2: Effect of changing BlobBoard generation parameters.
Figure 3: Blob detection-to-pose pipeline on a detected board. (a) Affine detections from the ASIFT scale space. (b) IRLS adaptation to anisotropic ellipses, magenta except the twelve pose inliers, which are coloured individually. (c) Those inliers canonicalized to log-polar patches, each bordered in its ellipse colour from (b). (d) Descriptors matched and spatially verified: the pattern is projected under the recovered pose, with reference blob centres in green at opacity proportional to their geometric consistency with it.
Figure 4: Descriptor architecture, adapted from HardNet [ Mishchuk et al.(2017)Mishchuk, Mishkin, Radenović, and Matas ] : log-polar input, kernels with padding , and column-wise max pooling of the rotation-equivariant feature map to a map that a dense layer maps to the -D descriptor. Kernel size, padding ( p ) and non-default stride ( s ) are shown above each layer; channel count below.
Figure 5: Experimental setup, one capture per marker type. Each type is mounted on a matched rigid panel carrying markers at three physical scales ( small, medium, large), observed while motion capture tracks the panel’s -DoF pose. Estimated board frames are overlaid; for BlobBoards the reference pattern is additionally projected under the recovered pose. Further captures are in supplementary fig. S1 .
Figure 6: Recall as a function of the error threshold, by marker size. Columns show small ( cm), medium ( cm), and large ( cm) markers; rows show translation and rotation error. A missed detection never enters the numerator, so each curve saturates at the corresponding probability of detection. Dashed lines mark recall .
Figure 7: Detection and pose accuracy over the capture floor plan. Concentric rings denote camera distance ( – m) and wedges denote nominal viewing-angle stations. Colour encodes (a) probability of detection and (b) median translation error. Grey wedges indicate visited stations with no detections.
Figure 8: Occlusion examples. Top: detected BlobBoards under physical occlusion, shown with recovered board projections and affine-adapted feature ellipses. Bottom: the synthetic occlusion mask applied to BlobBoards, AprilTag, and ArUco. Masks are defined in board coordinates and warped using the ground-truth homography, giving matched occlusion across methods and viewpoints.
Figure 9: Performance over the synthetic occlusion sweep for BlobBoards, AprilTag, and ArUco: (a) median rotation error, (b) median translation error, and (c) probability of detection. Medians are omitted when fewer than detections survive at an occlusion level, which is why the tag curves stop early.
| flips ( ) | AprilTag | ArUco | BlobBoards | BlobBoards+LRT |
|---|---|---|---|---|
| 3 | 21 | 0 | 0 | |
| 27 | 96 | 4 | 2 | |
| 94 | 35 | 41 | 31 | |
| total | 124 | 152 | 45 | 33 |
Table 1: Planar-pose flips (rotation error against motion-capture ground truth) by tripod distance. IPPE selects the lower-reprojection-error tag pose from four corners, whereas section 3.4 selects between BlobBoard branches using hundreds of verified correspondences. The final two columns use identical detections, correspondences, and calibration; the likelihood-ratio test (LRT) reduces the remaining flips from to .
| small | medium | large | |
| pattern extent (mm) | |||
| blobs per board (median) | |||
| shared across all sizes | |||
| minimum blob scale | mm | ||
| maximum diameter fraction | |||
| support cutoff | |||
Table S1: Pattern-generation settings for the evaluation boards ( small, medium, large). Each board is generated from its own random seed, so blob counts vary slightly within a size class; the ranges are – , – , and – .
| stage | parameter | value |
| detect | affine warps | IMAS- : identity, plus tilt at angles and tilt at angles |
| pre-upsampling | (scale space starts at octave ) | |
| DoG peak threshold | ||
| DoG edge threshold | ||
| cross-warp NMS | elliptical, Mahalanobis cutoff | |
| detections kept | unbounded (no response-rank cap) |
Table S2: Operational settings for the reported runs. The RANSAC seed is fixed and each board’s stream is derived from it, so the fit stage is bit-reproducible given the extracted features.
Figure S1: Experimental setup. BlobBoards, AprilTag, and ArUco are mounted on matched rigid panels containing markers at three physical scales: small, medium, and large. The panel is observed at varying distances and obliquities while its -DoF pose is tracked by motion capture. Estimated board coordinate frames are overlaid; for BlobBoards, the reference pattern is additionally projected under the recovered pose.
| system | frames | recovered | resid. mm | resid. deg |
|---|---|---|---|---|
| BlobBoards | ||||
| AprilTag | ||||
| ArUco |
Table S3: Hand–eye calibration diagnostics, one fit per marker system over all boards. Residual is the noise scale of the pose residual the staged linear fit reports, in millimetres and degrees; recovered counts calibration frames whose untracked camera pose was substituted. The camera is static; its rig was untracked in of the BlobBoard calibration frames, which use a substituted constant camera pose admitted only under an explicit constancy gate on the motion-capture reconstruction and the image background.
Figure S2: Recall as a function of the translation and rotation error threshold, pooled over all marker sizes. Every visible board constitutes an attempt; a missed detection never enters the numerator, so each curve saturates at the corresponding probability of detection. Dashed lines mark recall .
Figure S3: Unoccluded performance binned by obliquity and camera distance. Curves are per-bin medians, bands the interquartile range. Bins with fewer than five detections are omitted.
Figure S4: Pose accuracy over the capture floor plan at occlusion, in the same layout and encoding as the floor-plan figure of the main paper. Rings are tripod distance, wedges are rig stations; colour is (a) the fraction of the boards detected, and the per-station median over detected boards of (b) translation and (c) rotation error. Gray wedges are stations with zero detections. At this level the AprilTag and ArUco plans are already largely gray while the BlobBoards plan is intact — the separation between the two designs is established well before the heavier levels shown below.
Figure S5: Pose accuracy over the capture floor plan at occlusion, in the same layout and encoding as the floor-plan figure of the main paper. Rings are tripod distance, wedges are rig stations; colour is (a) the fraction of the boards detected, and the per-station median over detected boards of (b) translation and (c) rotation error. Gray wedges are stations with zero detections. Both tag plans are uniformly gray; the BlobBoards plan retains coverage across the capture floor at per-station errors in the range it holds unoccluded.
Figure S6: Pose accuracy over the capture floor plan at occlusion, in the same layout and encoding as the floor-plan figure of the main paper. Rings are tripod distance, wedges are rig stations; colour is (a) the fraction of the boards detected, and the per-station median over detected boards of (b) translation and (c) rotation error. Gray wedges are stations with zero detections. The AprilTag and ArUco plans are gray everywhere: at this level neither returns a pose anywhere on the capture floor. BlobBoards retain coverage over a reduced but still substantial set of stations, at per-station errors in the same range as the unoccluded case.
More in this topic
Object Pose Estimation
Object pose estimation recovers an object's 3D position and orientation from images, depth, point clouds, or contact sensing.
See all 187 papers