BlobBoards: Absolute-Pose Experiments and Descriptor Training Set

Version 1.3.0 — schema migration only; no data changed. Every versioned record is now in its current schema, read by BlobBoards.jl and VisualGeometryToolkitIO.jl 0.5. Each file was migrated once by VisualGeometryToolkitIO.vgio_migrate, which writes the new file through the current writer, reads it back and compares every value with the old file. Every other file is byte-identical to version 1.2.0. The captures are now a separate record. bb_absolute_pose_images_260527.tar.gz (2.3 GB, the raw captures of the 260527 session) is no longer in this record, so that the data downloads quickly: it is published as BlobBoards absolute-pose images (260527), doi:10.5281/zenodo.23965208, which supplements this record. Its tarball there is the one described below (without the two superseded camera files). BlobBoards.jl's bmvc2026/Artifacts.toml binds that tier to the images record and every other tier to this one. Versions 1.0.0 to 1.2.0 of this record keep their copy. Board definitions blobboards/board/v1 and v3 → blobboards/board/v5: the 30 BlobBoards panel boards in bb_absolute_pose_inputs (from v1), and the 50-board gallery in bb_assets and bb_descriptor_training_metadata (from v3). The stored identity hashes.config is carried over, so every board uid and file name is unchanged, and each migrated file is byte-identical to the same board in the 1.2.0 bb_boards_*_board5 tarballs. The AprilTag/ArUco board descriptors carry no versioned schema and are unchanged. Camera stores vgio/cameras/v3 → vgio/cameras/v4 (a lens's distortion stages listed in composition order under compose; the same cameras): the camera of record in bb_absolute_pose_inputs (260527/calibration/), the camera in bb_pipeline_figure, and the example still and clip, now bb_example_still_v4 and bb_example_video_v4. Removed from bb_absolute_pose_images_260527 (the tarball in the images record, doi:10.5281/zenodo.23965208): 260527/images/sony-fx3-iso100_tamron-20mm-f2.8-(set-to-f22).json and its _old.json. These are camera files in the retired vgio/cameras/v1 schema, which has no migration and no reader. They are superseded and were never read. The camera of record for the session is 260527/calibration/sony-fx3-iso100_tamron-20mm-f2.8-(set-to-f22).json in bb_absolute_pose_inputs, and its upstream calibration export ships beside it under provenance/. The captures themselves are unchanged. Not carried into 1.3.0: bb_example_still_v3, bb_example_video_v3 (replaced by the _v4 tarballs) and bb_example_boards_v3 (its boards as v5 are bb_boards_descriptor_training_wood_board5). They remain available in version 1.2.0. Re-packed tiers: bb_absolute_pose_inputs, bb_assets, bb_descriptor_training_metadata, bb_pipeline_figure, bb_example_still_v4, bb_example_video_v4. All other tarballs (the occlusion series, the training sources, the descriptor network and the six bb_boards_*_board5 board sets) are the version 1.2.0 files. Version 1.1.0 — schema migration only; no data changed. Every structured file is rewritten in the current schema through its owning package's own reader and writer; every other file is byte-identical to version 1.0.0. Board definitions: blobboards/board/v1 → blobboards/board/v3 (BlobBoards.jl 1.0): the stored identity hashes.config, and so every board uid and file name, is unchanged; blob centres, scales and saddle points are bit-identical; the ruler is stored as full-length ticks and the scale ladder as max_scale (the same printed boards). Camera calibrations, rigid transforms and unit sidecars: vgio/*/v1 → vgio/*/v2 (VisualGeometryIo.jl 1.0 migrate_v1, every record verified equal on read-back). Re-packed tiers: bb_assets, bb_descriptor_training_metadata, bb_absolute_pose_metadata, bb_absolute_pose_images_260527; all other tarballs are the version 1.0.0 files. New in 1.1.0: track_logpolar_circ_s96_iter4.tar.gz, the BlobBoards.jl package's descriptor network (previously hosted on Hugging Face; byte-identical tarball). BlobBoards.jl 1.0 / VisualGeometryIo.jl 1.0 refuse the v1 camera and transform files of version 1.0.0. Dataset accompanying BlobBoards: Robust Markers for Accurate Pose (BMVC 2026). It contains the absolute-pose evaluation data and the frozen descriptor training set, packaged as content-addressed Julia artifact tarballs that the bmvc2026 subproject of BlobBoards.jl (version 1.0, CauRegistry) consumes through its Artifacts.toml: each file below is one lazy artifact tier, verified by sha256 and git-tree-sha1 on download. Absolute-pose experiments (set 260527; mounted under absolute_pose_experiments/): bb_absolute_pose_metadata.tar.gz — the scientific layer: camera calibrations (vgio/cameras/v2, vgio/transforms/v2), motion-capture rig poses, hand-eye calibrations, the promoted BlobBoards pose snapshot and the fixed AprilTag/ArUco baseline detections for the unoccluded and every occlusion level, board definitions, provenance records. bb_absolute_pose_images_260527.tar.gz, bb_absolute_pose_val_260527.tar.gz — the raw calibrated captures (evaluation and validation frames) for the BlobBoards, AprilTag and ArUco rigs. bb_absolute_pose_occlusion_NN_260527.tar.gz (NN = 00…90) — the synthetic occlusion series: the evaluation frames with an occluder covering NN% of each board's pattern, one tarball per level. These are the frozen inputs of the published occlusion results for all three marker systems. bb_assets.tar.gz — the definitions of the 50 descriptor-training boards (wooden frames; the same boards as in bb_descriptor_training_metadata, which also serve as the identification gallery of the evaluation; definitions only, renders regenerate from them) and the descriptor networks of record (models/track_logpolar_circ_s96_iter4.onnx). Descriptor training set (mounted under descriptor_training/): bb_descriptor_training_sources.tar.gz — the 634 short indoor/outdoor photos and videos of the 50 training boards (disjoint from the 30 absolute-pose evaluation boards). bb_descriptor_training_metadata.tar.gz — the 50 training-board definitions, the bootstrap network chain (synthetic pretraining → three real-data iterations: configs, training logs, ONNX and PyTorch checkpoints), the track-building logs, and a README with the exact recipe that regenerates the training annotations from the sources. Board definitions, schema blobboards/board/v5 (added in 1.2.0; one tarball per physical board set, each board JSON migrated once from its earlier schema by VisualGeometryToolkitIO's vgio_migrate with BlobBoards loaded, its uid and stored hashes.config kept; print PDFs or PNG renders beside them where they exist). The board definitions inside the 1.1.0 tarballs above (board v1 and v3) are superseded by these and left unchanged: bb_boards_descriptor_training_wood_board5.tar.gz — the 50-board gallery, mounted on wooden frames (the boards of bb_assets, bb_descriptor_training_metadata and bb_example_boards_v3, from board v3). bb_boards_absolute_pose_panel_board5.tar.gz — the 30 boards of the absolute-pose evaluation, mounted on a panel (the boards of bb_absolute_pose_inputs, from board v1), with 4 PDFs and 26 PNG renders. bb_boards_mocap_mirrors_board5.tar.gz — the 50 boards of the 2026-09 motion-capture session, mounted on mirrors (from board v1), with their print PDFs. bb_boards_mocap_large_glass_board5.tar.gz — the large board of that session, mounted on glass (board e89; suspected bad), with its print PDF. bb_boards_print_series_aluminum_board5.tar.gz — six boards on aluminium, one per size (A5, A4, A3, A2, A1 and 1 m × 1 m), with their print PDFs (2 mm bleed) (from board v4). bb_boards_unknown_board5.tar.gz — two boards whose use is not recorded (7ba, from board v1; 79a, 1 m at 1200 dpi without frame, from board v2, with its PDF). Package artifacts (fetched by the package-root Artifacts.toml of BlobBoards.jl itself, not by bmvc2026): track_logpolar_circ_s96_iter4.tar.gz — the production descriptor network (track_logpolar_circ_s96_iter4.onnx) that the package's default configuration names through its artifact catalog. Licence: MIT. Code, evaluation drivers and the reproduction instructions live in the linked repository.