VisionQ / taxonomy /codebook.json
visionq-review's picture
VisionQ dataset release (part 13)
a32bc15 verified
Raw History Blame Contribute Delete
15.4 kB
{
"name": "VisionQ six-axis vocabulary (6 axes, 51 leaves)",
"kind": "root",
"children": [
{
"name": "Image Appearance",
"kind": "top_cluster",
"definition": "Whole-panel, crop-level, or panel-level low-level visual appearance.",
"evaluator_question": "Does the figure, crop, or panel have acceptable low-level visual appearance?",
"positive_scope": "Blur, noise, exposure, contrast, color, lighting, global artifact, realism, and style.",
"negative_scope": "Object shape errors, spatial relation errors, prompt mismatch, reference mismatch, point distribution, keypoint accuracy, mask accuracy, and latent attention-map behavior.",
"children": [
{
"name": "Blur",
"kind": "l1_cluster",
"definition": "Focus loss, smear, or blur visible at the panel or crop level.",
"children": []
},
{
"name": "Noise",
"kind": "l1_cluster",
"definition": "Random speckle, grain, sensor-like corruption, or noisy visual artifacts.",
"children": []
},
{
"name": "Exposure",
"kind": "l1_cluster",
"definition": "Over-exposure, under-exposure, saturation clipping, or visibility loss from brightness.",
"children": []
},
{
"name": "Contrast",
"kind": "l1_cluster",
"definition": "Insufficient or excessive tonal separation that harms visible interpretation.",
"children": []
},
{
"name": "Color",
"kind": "l1_cluster",
"definition": "Color cast, color inconsistency, or visually implausible color independent of prompt/reference criteria.",
"children": []
},
{
"name": "Lighting",
"kind": "l1_cluster",
"definition": "Illumination, shadow, reflection, or lighting quality visible in the output.",
"children": []
},
{
"name": "Global Artifact",
"kind": "l1_cluster",
"definition": "Panel-wide artifact not better explained by object form, layout, relation, prompt, or reference match.",
"children": []
},
{
"name": "Realism",
"kind": "l1_cluster",
"definition": "Overall perceptual plausibility or photorealism when not tied to a specific object, prompt, or reference.",
"children": []
},
{
"name": "Style",
"kind": "l1_cluster",
"definition": "Global visual style or aesthetic quality when judged directly rather than as reference or prompt match.",
"children": []
}
]
},
{
"name": "Object Form",
"kind": "top_cluster",
"definition": "Internal quality of visible objects, regions, bodies, faces, surfaces, and masks.",
"evaluator_question": "Are visible objects or regions internally well-formed?",
"positive_scope": "Shape, boundary, texture, surface, material, completeness, pose, face, anatomy, detail, and segmentation boundary quality.",
"negative_scope": "Whole-scene arrangement, pairwise relation, explicit prompt matching, and visible reference or ground-truth matching.",
"children": [
{
"name": "Shape",
"kind": "l1_cluster",
"definition": "Object geometry, silhouette, deformation, or structural shape correctness.",
"children": []
},
{
"name": "Boundary",
"kind": "l1_cluster",
"definition": "Object or region edge quality, contour sharpness, silhouette completeness, or boundary artifacts.",
"children": []
},
{
"name": "Texture",
"kind": "l1_cluster",
"definition": "Local texture quality, texture continuity, or texture corruption on visible regions.",
"children": []
},
{
"name": "Surface",
"kind": "l1_cluster",
"definition": "Surface smoothness, topology, wrinkles, holes, or visible surface defects.",
"children": []
},
{
"name": "Material",
"kind": "l1_cluster",
"definition": "Material appearance or material property quality of an object or region.",
"children": []
},
{
"name": "Completeness",
"kind": "l1_cluster",
"definition": "Missing, truncated, duplicated, or incomplete object parts.",
"children": []
},
{
"name": "Pose",
"kind": "l1_cluster",
"definition": "Pose plausibility of a single visible object, body, or articulated entity when not judged against ground truth.",
"children": []
},
{
"name": "Face",
"kind": "l1_cluster",
"definition": "Face-specific form quality including expression quality when not prompt- or reference-driven.",
"children": []
},
{
"name": "Anatomy",
"kind": "l1_cluster",
"definition": "Body, limb, hand, joint, or anatomical plausibility of one entity.",
"children": []
},
{
"name": "Detail",
"kind": "l1_cluster",
"definition": "Fine detail, thin structures, small parts, or local feature preservation.",
"children": []
},
{
"name": "Segmentation",
"kind": "l1_cluster",
"definition": "Mask or segmentation visual quality when judged as object-region form rather than GT match.",
"children": []
}
]
},
{
"name": "Scene Layout",
"kind": "top_cluster",
"definition": "Scene-level spatial organization and physical arrangement.",
"evaluator_question": "Are scene-level spatial properties correct?",
"positive_scope": "Composition, depth, scale, view, spatial distribution, occlusion layout, camera geometry, and physical plausibility.",
"negative_scope": "Object-internal defects, exact pairwise relation claims, explicit reference matching, and prompt-specific layout instructions.",
"children": [
{
"name": "Composition",
"kind": "l1_cluster",
"definition": "Overall arrangement of regions, objects, foreground, background, or panel contents.",
"children": []
},
{
"name": "Depth",
"kind": "l1_cluster",
"definition": "Depth ordering, 3D arrangement, or perceived scene depth.",
"children": []
},
{
"name": "Scale",
"kind": "l1_cluster",
"definition": "Relative or absolute size plausibility at scene level.",
"children": []
},
{
"name": "View",
"kind": "l1_cluster",
"definition": "Viewpoint, novel-view layout, or multi-view scene-level consistency.",
"children": []
},
{
"name": "Spatial Distribution",
"kind": "l1_cluster",
"definition": "Distribution, spacing, clumping, or spread of visible scene elements or point samples.",
"children": []
},
{
"name": "Occlusion Layout",
"kind": "l1_cluster",
"definition": "Occlusion ordering or impossible scene-level overlap patterns.",
"children": []
},
{
"name": "Camera Geometry",
"kind": "l1_cluster",
"definition": "Camera perspective, projection geometry, or viewpoint geometry visible in the figure.",
"children": []
},
{
"name": "Physical Plausibility",
"kind": "l1_cluster",
"definition": "Scene-level physical plausibility such as floating objects, floor contact at scene scale, or impossible arrangements.",
"children": []
}
]
},
{
"name": "Relation",
"kind": "top_cluster",
"definition": "Correctness of relationships between visible entities, parts, attributes, or correspondence points.",
"evaluator_question": "Are relationships between visible entities correct?",
"positive_scope": "Contact, correspondence, attribute binding, spatial relation, interaction, and part-whole relation.",
"negative_scope": "Single-entity pose plausibility, GT/keypoint accuracy, bounding-box accuracy, segmentation quality, and protocol robustness.",
"children": [
{
"name": "Contact",
"kind": "l1_cluster",
"definition": "Whether two visible entities or parts touch, connect, intersect, or maintain contact correctly.",
"children": []
},
{
"name": "Correspondence",
"kind": "l1_cluster",
"definition": "Visible correspondence between entities, parts, points, views, or panels.",
"children": []
},
{
"name": "Attribute Binding",
"kind": "l1_cluster",
"definition": "Correct binding of visible attributes to the right entity or part.",
"children": []
},
{
"name": "Spatial Relation",
"kind": "l1_cluster",
"definition": "Pairwise relation such as left/right, above/below, inside/outside, near/far, or in-front/behind.",
"children": []
},
{
"name": "Interaction",
"kind": "l1_cluster",
"definition": "Visible interaction between people, objects, agents, tools, or scene elements.",
"children": []
},
{
"name": "Part-Whole Relation",
"kind": "l1_cluster",
"definition": "Relationship between an object and its visible parts when the relation itself is the criterion.",
"children": []
}
]
},
{
"name": "Reference Match",
"kind": "top_cluster",
"definition": "Fidelity to a visible source, target, ground truth, before-after panel, or comparison reference.",
"evaluator_question": "Does the output match a visible source, target, ground truth, or comparison panel?",
"positive_scope": "Identity, background, reconstruction, edit preservation, geometry match, style match, mask match, landmark match, and segmentation match.",
"negative_scope": "Prompt-only matching, absolute image appearance with no reference, single-object form with no visible reference, and protocol robustness.",
"children": [
{
"name": "Identity",
"kind": "l1_cluster",
"definition": "Identity preservation or identity match against a visible source/reference.",
"children": []
},
{
"name": "Background",
"kind": "l1_cluster",
"definition": "Background preservation or match to visible source/reference.",
"children": []
},
{
"name": "Reconstruction",
"kind": "l1_cluster",
"definition": "Reconstruction fidelity against target, source, or ground truth.",
"children": []
},
{
"name": "Edit Preservation",
"kind": "l1_cluster",
"definition": "Preservation of non-edited content in before/after or edit-reference setups.",
"children": []
},
{
"name": "Geometry Match",
"kind": "l1_cluster",
"definition": "Geometric fidelity to visible source, target, annotation, or ground truth.",
"children": []
},
{
"name": "Style Match",
"kind": "l1_cluster",
"definition": "Style fidelity when a visible reference style is the criterion.",
"children": []
},
{
"name": "Mask Match",
"kind": "l1_cluster",
"definition": "Mask fidelity against visible mask, annotation, or ground truth.",
"children": []
},
{
"name": "Landmark Match",
"kind": "l1_cluster",
"definition": "Keypoint, landmark, or pose-coordinate match against visible ground truth or annotation.",
"children": []
},
{
"name": "Segmentation Match",
"kind": "l1_cluster",
"definition": "Segmentation output fidelity against visible ground-truth segmentation.",
"children": []
}
]
},
{
"name": "Prompt Match",
"kind": "top_cluster",
"definition": "Match to a text prompt, caption, instruction, requested content, or requested attribute.",
"evaluator_question": "Does the output match the text prompt, caption, or instruction?",
"positive_scope": "Object presence, attribute match, spatial instruction, semantic match, text legibility, count match, action match, and expression match.",
"negative_scope": "Lighting, expression, physical plausibility, or object quality when not specified by prompt/caption/instruction.",
"children": [
{
"name": "Object Presence",
"kind": "l1_cluster",
"definition": "Requested object, entity, category, or scene element is present or absent as specified.",
"children": []
},
{
"name": "Attribute Match",
"kind": "l1_cluster",
"definition": "Requested color, material, style, state, or other attribute is correctly assigned.",
"children": []
},
{
"name": "Spatial Instruction",
"kind": "l1_cluster",
"definition": "Prompt-specified spatial arrangement or relation is satisfied.",
"children": []
},
{
"name": "Semantic Match",
"kind": "l1_cluster",
"definition": "Overall semantic content matches the caption, prompt, or instruction.",
"children": []
},
{
"name": "Text Legibility",
"kind": "l1_cluster",
"definition": "Requested text, glyphs, symbols, or readable writing are rendered correctly.",
"children": []
},
{
"name": "Count Match",
"kind": "l1_cluster",
"definition": "Requested number of objects or entities is correct.",
"children": []
},
{
"name": "Action Match",
"kind": "l1_cluster",
"definition": "Requested action, activity, or event is depicted correctly.",
"children": []
},
{
"name": "Expression Match",
"kind": "l1_cluster",
"definition": "Requested facial expression, emotion, or pose state is depicted correctly.",
"children": []
}
]
}
],
"sidecar": {
"name": "Sidecar / Rejected",
"kind": "sidecar",
"definition": "Research concepts that should not be forced into the main qualitative figure tree.",
"reason_codes": [
"temporal_out_of_scope",
"method_or_nonvisual_property",
"protocol_robustness",
"latent_internal_metric",
"dataset_training_property",
"metric_calibration",
"multi_sample_required",
"not_visible_in_figure",
"requires_external_knowledge"
],
"children": [
{
"name": "Temporal Behavior",
"kind": "sidecar_cluster",
"definition": "Claims requiring video playback, sequence behavior, flicker, jitter, or temporal dynamics."
},
{
"name": "Multi-Sample Diversity",
"kind": "sidecar_cluster",
"definition": "Claims requiring multiple generated samples or distribution-level diversity evidence."
},
{
"name": "Method Control",
"kind": "sidecar_cluster",
"definition": "Claims about controllability or intervention behavior not directly judged from a single qualitative figure."
},
{
"name": "Protocol Robustness",
"kind": "sidecar_cluster",
"definition": "Claims about robustness to perturbations, occlusion protocols, or test-time conditions."
},
{
"name": "Latent Representation",
"kind": "sidecar_cluster",
"definition": "Claims about attention maps, internal features, embeddings, or hidden model states when not used as visible figure output."
},
{
"name": "Dataset Coverage",
"kind": "sidecar_cluster",
"definition": "Claims about dataset distribution, coverage, domain breadth, or benchmark composition."
},
{
"name": "Training Memorization",
"kind": "sidecar_cluster",
"definition": "Claims about memorization, training data leakage, or learned representations."
},
{
"name": "Metric Calibration",
"kind": "sidecar_cluster",
"definition": "Claims about numerical metrics, calibration, FID, LPIPS, IoU, or other score behavior."
},
{
"name": "Nonvisual Method Property",
"kind": "sidecar_cluster",
"definition": "Claims about speed, efficiency, architecture, optimization, or generalization without direct visual figure evidence."
}
]
}
}