PaperlayerView exact revision on arXiv

Source-linked AI summary

VisionQ: VLM-as-a-Judge Taxonomy, Dataset and Benchmark for Qualitative Analysis in Computer Vision

Vu Dinh Xuan, Duc-Hai Nguyen, Minh-Dung Dao, Vu Quynh Giao, Quang Hong Nguyen, Binh-Son Hua, Barry O'Sullivan, David Murphy, Hoang D. Nguyen

arXiv:2610.00666v1cs.CVcs.AI

TL;DR

Existing VLM-judge benchmarks evaluate whether a model selects a preferred image, not whether it identifies the visual criterion behind that choice. VisionQ builds a criterion-conditioned benchmark and taxonomy from computer-vision papers, and its evaluation shows that judge accuracy varies by criterion.

  • Problem

    Existing evaluations score preferred-image verdicts without checking whether a judge identifies the visual criterion or grounds its rationale in figure evidence.

    Source: §1 · pp. 1–2

  • Method

    VisionQ pairs author-stated visual criteria with comparison figures, using a six-axis taxonomy and questions that credit only the output best on the named criterion.

    Source: §1 · p. 2, §2.2 · p. 4

  • Results

    Criterion-specific accuracy varies: one of the strongest judges reaches 64.1% on Image Appearance but 45.2% on Relation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The taxonomy provides a vocabulary for criterion-level and rationale evaluation that may generalize to other visual comparison settings.

    Source: §6 · p. 14

    The source corpus is limited to CVPR and ICCV, whose visual-domain representation may differ from the broader CV literature.

    Source: §6 · p. 14

Abstract

from arXiv · show

Qualitative comparison figures are central evidence in computer vision papers, and vision-language models (VLMs) are increasingly used to judge them. Yet existing benchmarks score only scalar quality or overall preference, so a judge can be rewarded for picking the preferred image for the wrong visual reason. We introduce VisionQ, the first benchmark built from peer-reviewed CV comparison figures that grounds every judgment in a named visual criterion: each question states the criterion, and a judge is credited only when it selects the output the authors identify as best on that criterion. We call this task criterion-conditioned visual discrimination. VisionQ comprises (1) a corpus of 1,409 CVPR and ICCV papers with 1,800+ validated comparison figures and 3,911 hand-annotated data points linking method crops to author-stated visual claims; (2) a six-axis, 51-leaf taxonomy of the visual criteria behind qualitative judgment; (3) a criterion-conditioned evaluation protocol that hides method names, captions, and paper identity and reports accuracy per criterion; and (4) VisionQ-Judge, a DPO-tuned Gemma-4-E4B judge trained on symmetric evidence pairs, which reduces last-option predictions by 7.0pp and improves accuracy by 2.5pp on a held-out test set. Evaluating 20 open- and closed-source VLM judges, we find that the strongest reach only 63.1% accuracy (chance 32.2%) and that reliability varies sharply across criteria. Code: https://github.com/ReML-AI/visionq. Data: https://huggingface.co/datasets/visionq-anon-2026/VisionQ-1k.

1 Introduction

VisionQ makes qualitative figure judgment criterion-conditioned, addressing the gap between selecting a preferred image and identifying the visual reason for that preference. It pairs a taxonomy-grounded benchmark with a judge trained on symmetric evidence pairs and a study of criterion-specific performance.

1 Introduction

  • Existing verdict-only benchmarks credit the preferred-image choice even when its stated visual evidence is unsupported, limiting trust in rationales and diagnosis of failures.This gap matters for computer-vision comparison figures, which are central evidence for method evaluation.Source: §1 · pp. 1–2
  • VisionQ names an author-stated visual criterion in each question and credits only the output identified as best on that criterion.The protocol hides method names, captions, and paper identity, and reports accuracy per criterion.Source: §1 · p. 2
  • One of the strongest judges reaches 64.1% on Image Appearance but only 45.2% on Relation, illustrating criterion-specific variation.The study profiles 20 open- and closed-source VLM judges for overall and per-criterion accuracy.Source: §1 · p. 2
  • The benchmark draws on 1,409 peer-reviewed CV papers, 1,800+ validated comparison figures, and 3,911 hand-annotated data points.Its six-axis taxonomy is grounded in corpus-driven claim analysis and prior benchmark dimensions.Source: §1 · p. 2
  • VisionQ-Judge uses DPO fine-tuning on symmetric evidence pairs, reducing last-option predictions by 7.0 pp and improving preference accuracy by +2.5 pp.The training set comprises 4,524 VisionQ-MCQ questions spanning 513 papers.Source: §1 · p. 2

2 Taxonomy Construction

VisionQ builds a six-axis, 51-leaf visual-criteria taxonomy from prior evaluation dimensions and claims extracted from CV papers. It validates the taxonomy’s coverage and uses it to construct filtered, criterion-conditioned multiple-choice questions.

2 Taxonomy Construction

  • The taxonomy combines prior visual-evaluation dimensions with 9,228 grounded claims from 1,409 papers, requiring criteria to be reviewer-like, figure-observable, and specific enough for reproducible annotation.It is built to guide annotation and benchmark synthesis, while explicitly excluding non-visual claims.Source: §2 · p. 2

2.1 Two-Stage Construction

  • The two-stage construction seeds six axes from prior work, canonicalizes 447 recurring claim atoms, then organizes them into 51 leaves with human-reviewed LLM-assisted drafts.The hierarchy procedure clusters candidate leaves and groups them under mid-level categories and six top-level axes.Source: §2.1 · pp. 3–4

2.2 Final Codebook

  • The final codebook groups six axes across image-, scene-, and cross-image criteria, assigning overlapping claims by a fixed priority order and excluding non-figure-observable claims.The priority convention was not tested against alternative orderings; out-of-scope claims receive reason codes and are excluded from axis assignment and the benchmark.Source: §2.2 · pp. 4–5

2.3 Validation

  • In the coverage check, 1,426 of 1,486 assignments (96.0%) fit the taxonomy; task types also show distinct leaf distributions that support the 51-leaf vocabulary.Object Form accounts for 43.4% of assignments, while Scene Layout accounts for 3.6%; reconstruction 3D, generative 2D, and segmentation/detection have different claim distributions.Source: §2.3 · p. 5

2.4 From Taxonomy to Benchmark

  • VisionQ-Bench questions pair candidate crops with leaf-level criteria, showing only a composite image and question stem; filtering leaves 309 eligible items from 332 synthesized MCQs.The filter removes invalid comparisons involving duplicate method labels, context panels, or single-method-family ablations.Source: §2.4 · p. 5, Figure 3 · p. 6, Table 1 · p. 6

3 Dataset

VisionQ builds a corpus of peer-reviewed computer-vision comparison figures and author-stated claims, then uses them to analyze dataset coverage and selective comparison practices.

3 Dataset

  • The corpus serves both taxonomy construction across 1,409 papers and a figure funnel that supplies the benchmark pool.The taxonomy claims are drawn independently of whether a paper later contributes figures to the benchmark.Source: §3 · p. 6, §3.1 · p. 7

3.1 Source Corpus and Figure Selection

  • From 3,651 extracted candidates, caption and text cues identify 1,800+ validated qualitative-comparison figures.Figures require method outputs alongside baselines or references in a shared panel that supports a visual preference judgment.Source: §3.1 · pp. 6–7

3.2 Hand-Curated Subset: Annotation Funnel and Final Release

  • Annotation and quality review yield 3,911 data points across 1,486 figures and a 4,524-question release spanning 707 figures from 513 papers.Out-of-scope data points are excluded from the 3,773 in-scope points; records failing crop, method-identification, claim-recovery, or review checks are dropped.Source: §3.2 · p. 7

3.3 Benchmark Item Construction

  • Each benchmark item links candidate outputs, method identifiers, an author-stated claim, and assigned taxonomy axes; items without a single winner are discarded.One data point can produce multiple question variants, and all questions from that data point stay on the same side of each train/test split.Source: §3.3 · pp. 7–8

3.4 Coverage Analysis and Field Statistics

  • 790 papers cover all 51 observed taxonomy leaves across 20 task types, while mean data points per covered paper range from 3.1 in motion generation to 6.5 in novel view synthesis.No task type exceeds 18.0% of in-scope data points; the paper attributes depth differences to domain-specific comparison-panel practices.Source: §3.4 · p. 8

3.5 Selective Comparison Analysis

  • 59% of 789 selective-comparison candidates test exactly one of their field’s top 10 criteria; the mean untested count is 8.73.The authors frame selective comparison as a coverage diagnostic, not an assessment of paper merit or correctness; they leave testing omitted criteria in a 200-paper sample to future work.Source: §3.5 · p. 8, Figure 4 · p. 9

4 Benchmark Protocols and Model Evaluation

VisionQ evaluates criterion-conditioned visual judgments with anonymized comparison questions and reports performance across models and criteria. The strongest judges reach 0.631 accuracy, while small axis samples limit some per-axis conclusions.

Evaluation protocol

  • Each question shows a lettered composite image and leaf-level criterion while hiding paper identity, method names, captions, and author claims.The protocol isolates visual judgment from method-name priors and textual leakage.Source: §4 · p. 9

Question sets

  • VisionQ-Bench is a fixed 309-question evaluation set, while the 4,524-question VisionQ-MCQ release supports VisionQ-Judge training and paper-disjoint evaluation.The two question sets come from the same annotated figure pool.Source: §4 · p. 9

Eligible evaluation slice

  • The eligible slice spans six taxonomy axes, but Scene Layout (n = 8) and Prompt Match (n = 22) are too small for reliable per-axis conclusions.Their sizes reflect annotation yield under quality gating, not a sampling choice.Source: §4 · pp. 9–10

Metrics

  • The study reports raw accuracy, counts invalid or failed responses as incorrect, and also defines a chance-normalized score for the released dashboard.All paper comparisons use raw accuracy; the normalized score is 0 at chance, 1 at perfect accuracy, and negative below chance.Source: §4 · p. 10

Model comparison

  • 0.631 accuracy is the highest point estimate, against chance level 0.322; VisionQ-Judge reaches 0.482 versus 0.450 for its base model.The VisionQ-Bench result is held out from VisionQ-Judge training; its 4B model is within 1 pp of Gemma-4-26B-A4B, Mistral Medium 3.5, and Haiku 4.5.Source: §4 · p. 10

5 VisionQ-Judge: Specialising a VLM via DPO

VisionQ-Judge is built by constructing evidence-grounded DPO pairs and fine-tuning Gemma-4-E4B to address positional bias and the domain gap. Held-out results show a modest overall accuracy gain, a reduction in last-option bias, and sharply varying per-leaf outcomes.

5.1 LLM-Free DPO Dataset Construction

  • DPO pairs were built without LLM calls from structured author claims, visual-attribute descriptions, winner labels, and per-method crops.The metadata yielded 4,524 questions spanning 513 papers and 46 taxonomy leaves.Source: §5.1 · p. 10, Figure 6 · p. 11

5.2 Symmetric Evidence Pairs

  • Chosen and rejected responses share a one-sentence evidence string and differ only in the referenced answer letter, preventing response length or style from serving as a shortcut.Regex substitution succeeded in 90.4% of cases; otherwise, pairs used bare letters. The training set was capped at 300 questions per leaf and split by paper, yielding 2,867 training and 717 test questions; position-swap augmentation and pairing each correct answer with each wrong option produced 9,525 pairs.Source: §5.2 · pp. 10–11, Figure 7 · p. 11

5.3 Results

  • Overall accuracy rose from 49.9% to 52.4% (+2.5 pp) on the 717-question held-out test set.The gains were +3.7 pp for 2-choice and +3.2 pp for 3-choice questions; 4-choice accuracy changed −0.5 pp (one question).Source: Table 2 · p. 12, §5.3 · p. 12
  • Last-letter answers fell from 54.1% to 47.1% (−7.0 pp), toward the 38.9% gold share; accuracy on non-last-answer questions rose from 36.5% to 43.8% (+7.3 pp).The overall accuracy gain was modest (+2.5 pp, 95% CI −1.9 to +7.0), while the last-letter-share reduction had a 95% CI of −10.9 to −3.2; the 3-choice C-bias remained at 48.6% versus 35.0% gold. Position-swap augmentation was used in the same run, so the change is not attributed to either component alone.Source: §5.3 · p. 12, Table 3 · p. 12, Figure 8 · p. 13

5.4 Per-Leaf Analysis

  • The conclusions are bounded by small per-leaf test support, evaluation of one base model, and programmatically extracted evidence strings that do not cover the full visual reasoning space.Nineteen of 38 test leaves came from a single paper; only Semantic Match had a multi-paper change with a paper-clustered 95% interval excluding zero.Source: §5.4 · p. 13
  • Largest gains were Lighting (+28.6 pp, n = 7), Anatomy (+25.0 pp, n = 8), and Blur (+21.1 pp, n = 19).Attribute Binding (+20.0 pp, n = 10) and Landmark Fidelity (+18.2 pp, n = 11) also improved; Lighting, Anatomy, and Attribute Binding each came from one test paper. The authors suggest holistic, immediately legible differences may be easier to correct at 512 px tile resolution.Source: §5.4 · p. 13, Figure 9 · p. 14
  • Largest declines were Noise (−36.4 pp, n = 11), Semantic Match (−27.6 pp, n = 29), Segmentation (−14.3 pp, n = 14), Identity Preservation (−14.3 pp, n = 14), and Texture (−5.3 pp, n = 38).The authors suggest ceiling disruption for already high-accuracy Noise and Semantic Match leaves, and insufficient targeting of fine-grained identity and texture signals within 512 px tiles.Source: §5.4 · p. 13, Figure 9 · p. 14

6 Discussion and Limitations

VisionQ establishes infrastructure for evaluating reviewer-like qualitative figure judgment and releases resources intended to support evaluation beyond verdict accuracy. Its current scope is bounded by the source venues and author-stated ground truth, among other limitations.

6 Discussion and Limitations

  • The corpus draws only from CVPR and ICCV, which over-represent some visual domains, and its author-stated claims may be imprecise or biased and are not checked against human perceptual ground truth.Axis sizes also follow annotation yield: Object Form dominates while Scene Layout is sparse, making results on small axes indicative only.Source: §6 · p. 14
  • Possible pretraining exposure is mitigated by hiding paper identity, captions, and method names and recomposing figures, but contamination is not directly tested or ruled out.The reported exploratory comparison is not a direct test because every evaluated model was released after 2024.Source: §6 · p. 14
  • The benchmark lacks a human-accuracy baseline, and VisionQ-Judge’s DPO results concern one base model rather than multiple model families.The best score should therefore be read against chance, not as human-level performance.Source: §6 · p. 14
  • Static comparison figures exclude temporal claims about video quality or motion consistency, leaving methods focused primarily on video underrepresented in the hand-curated subset.The benchmark scores panel choices under a named criterion, not free-text rationales or whether a rationale causally explains a verdict.Source: §6 · p. 14
  • VisionQ releases corpus metadata, taxonomy, questions, predictions, and code as infrastructure for developing evaluation methods beyond verdict accuracy.Its taxonomy may also generalize to other visual comparison settings and rationale evaluation built on it.Source: §6 · pp. 14–15

A.1 Image Quality Assessment and Low-Level Visual Benchmarks

Prior benchmarks assess image quality, preference, or structured visual reasoning, but do not reproduce the reviewer-like task of judging qualitative figures from computer-vision papers. VisionQ conditions judgments on named criteria and grounds them in author-stated claims.

A.1 Image Quality Assessment and Low-Level Visual Benchmarks

  • Image-quality benchmarks assess low-level perception or scalar quality, while VisionQ instead compares outputs from paper figures against a named visual criterion.The cited work covers quality description, comparison, and rating; discrete human ratings; generated-image evaluation; and ranked comparisons with rationales.Source: §A.1 · p. 18

A.2 VLM-as-Judge and Preference Evaluation

  • Existing VLM-as-judge benchmarks measure verdict accuracy or calibration, so a correct choice supported by incorrect visual evidence can still count as correct.VisionQ conditions verdicts on a named criterion and retains author claims needed to evaluate justifications.Source: §A.2 · p. 18

A.3 Visual Reasoning, Text-Image Alignment, and Compositional Benchmarks

  • Visual reasoning and alignment benchmarks test curated prompts or generated content rather than qualitative figures from peer-reviewed CV papers grounded in author claims.VisionQ is constructed from that source for reviewer-like judgment; related benchmarks include per-axis analysis and subtle-difference tests.Source: §A.3 · p. 19

A.4 Qualitative Evaluation in Computer-Vision Papers

VisionQ addresses the lack of a benchmark for reviewer-like qualitative comparison in computer-vision papers by tying each judgment to an author-claimed, named visual criterion. Its taxonomy codebook defines six axes and their boundaries for assigning visual claims.

A.4 Qualitative Evaluation in Computer-Vision Papers

  • Prior qualitative evaluation relies on human inspection or crowd-sourced preferences without requiring specific criteria or grounding; VisionQ supplies figure-level ground truth from author claims and conditions each judgment on a named criterion.The accompanying codebook specifies each axis’s definition, evaluator question, positive and negative scope, and 51 leaf definitions.Source: §A.4 · p. 19, B Full Taxonomy Codebook · p. 19

B Full Taxonomy Codebook

  • The taxonomy codebook defines axis boundaries and evaluator questions, and Table 4 groups its 51 leaves by axis, with a separate Reference Fidelity split.Per-leaf questions, examples, and intermediate groupings are omitted except for the Reference Fidelity split.Source: B Full Taxonomy Codebook · p. 19, Table 4 · p. 21

B.1 Image Appearance

  • Image Appearance covers low-level whole-panel or crop qualities such as blur, noise, color, realism, and style, but excludes object-shape, relation, prompt, and reference errors.Scale or view consistency belongs to Scene Layout, and multi-sample diversity is out of scope; malformed anatomy belongs to Object Form.Source: §B.1 · pp. 19–20

B.2 Object Form

  • Object Form concerns the internal quality of visible objects or regions, including shape, texture, completeness, pose, anatomy, detail, and segmentation boundaries.It excludes whole-scene arrangement, pairwise relations, prompt matching, and visible-reference matching.Source: §B.2 · p. 20

B.3 Scene Layout

  • Scene Layout covers scene-level spatial organization and physical arrangement, including composition, depth, scale, view, occlusion, camera geometry, and physical plausibility.It excludes object-internal defects, exact pairwise relations, reference matching, and prompt-specific layout instructions.Source: §B.3 · p. 20

B.4 Relation

  • Relation covers correctness between visible entities, parts, attributes, or correspondence points, including contact, attribute binding, spatial relations, and interaction.A single malformed or missing part belongs to Object Form, not Relation.Source: §B.4 · p. 20

B.5 Reference Fidelity

  • Reference Fidelity asks whether an output preserves or matches a visible reference, such as a source image, target, ground truth, mask, or landmark.Pure method comparison without a visible reference is assigned to the axis matching the actual visual criterion.Source: §B.5 · p. 20

B.6 Prompt Match

  • Prompt Match concerns whether an output satisfies text prompts, captions, or instructions, including requested objects, attributes, layouts, text, counts, actions, and expressions.The codebook distinguishes prompt-specified qualities from lighting, expression, plausibility, or object quality not specified by the prompt.Source: §B.6 · p. 20

C Full Benchmark Profile

The full benchmark profile reports accuracy for 20 models on an eligible slice of 309 questions. It presents overall and per-axis results with confidence intervals, while noting that paper-clustered intervals are wider because questions within a paper are not independent.

C Full Benchmark Profile

  • The full experiment evaluates 20 models on 309 eligible questions, with Figure 10 showing headline accuracy and treating unparsed responses as incorrect.VisionQ-Judge is evaluated on its held-out VisionQ-Bench test split, with its base model shown for reference.Source: C Full Benchmark Profile · p. 22, Figure 10 · p. 23
  • Figure 11 reports per-axis accuracy with Wilson 95% confidence intervals for axes with more than 30 samples, sorting models by overall accuracy.The dashboard also provides per-axis norm_score, parameter-scaling plots for models with disclosed sizes, and per-model strongest and weakest leaves.Source: C Full Benchmark Profile · p. 22, Figure 11 · p. 23
  • Table 5 reports per-axis accuracy with paper-clustered 95% intervals, which are wider than the figures’ Wilson intervals because questions from the same paper are not independent.The dashboard supplies additional per-axis and per-model breakdowns.Source: C Full Benchmark Profile · pp. 22–23

D.1 LLM-Free DPO Dataset Construction

VisionQ’s DPO questions are drawn from validated semantic records that link author-stated claims and taxonomy assignments to method crops. The construction filters for genuine comparisons, identifies a claim-supported winner, and presents method crops in tiled form.

D.1 LLM-Free DPO Dataset Construction

  • Each retained record includes a claim, taxonomy assignment, method-crop bounding boxes, and winner-evidence provenance.The candidate pool requires proposed or validated status, a main-tree gate decision, and winner evidence.Source: §D.1 · pp. 23–24
  • Questions use pair_compare, reference_pair, or joint_pair modes, excluding single_panel records without a comparison partner.Source: §D.1 · p. 24
  • 4,524 usable questions cover 46 of the 51 taxonomy leaves after winner matching and utilization-based sampling.Records without an identifiable winner bounding box are discarded.Source: §D.1 · p. 24
  • Each question compares 2–4 methods, excludes reference-matched distractors, and assigns the winner a reproducible random letter position.Only one preferred bounding box is retained per method slug.Source: §D.1 · p. 24

D.2 Image Preprocessing and Tiling

  • Method crops are composited into one labelled tile, matching inference-time presentation and avoiding a multi-image DPO tensor-shape failure.Crops are square-padded, resized to at most 512 px, arranged in a grid, and labelled A–D.Source: §D.2 · p. 24
  • The source corpus spans venues and years; 2024 contains CVPR papers only because ICCV is held in odd years.Source: Figure 12 · p. 25

D.3 Training Configuration

VisionQ-Judge is trained with a one-epoch DPO configuration using Gemma-4-E4B and a language-decoder LoRA adapter. A paper-disjoint, leaf-stratified split and position-swap augmentation define the training data.

D.3 Training Configuration

  • Gemma-4-E4B-it is trained in bfloat16 on one 46 GB GPU, with per-sample batch size 1 and gradient accumulation 8.The model has 4.5B effective parameters and 8B total parameters with embeddings.Source: §D.3 · p. 25
  • The decoder-only LoRA adapter uses rank 16, α = 32, and dropout 0.05, while excluding the vision-tower wrapper.Target modules are the language decoder’s q, k, v, o, gate, up, and down projections.Source: §D.3 · p. 25
  • DPO uses β = 0.3, learning rate 5×10−6, 5% warmup, max length 1,024 tokens, and a 512 px image limit for one epoch.The learning rate decays linearly.Source: §D.3 · p. 25
  • The paper-disjoint, leaf-stratified split yields 2,867 training and 717 test questions, then position swaps expand training to 5,729 questions and 9,525 pairs.Each leaf is capped at 300 questions before splitting; whole-paper assignment results in a realised test share of 20%.Source: §D.3 · p. 25

E Coverage Analysis Supplement

The appendix provides coverage statistics and field-evaluation visualizations as supporting material for the paper’s coverage, taxonomy, and selective-comparison analyses.

E Coverage Analysis Supplement

E.1 Source Corpus Distribution

The source corpus spans many computer-vision task types, with comparison figures and evaluation coverage distributed unevenly across fields. The appendix documents field-specific criterion profiles and describes how eligible papers were sampled for selective-comparison assessment.

E.1 Source Corpus Distribution

  • 97.4% of papers and 95.0% of annotated figures contain a group with ≥2 images, across 9,382 groups with a median of 4 images.Among groups, 88.0% contain at least two images, 76.3% at least three, and 64.2% at least four; the comparison-group statistics use a population distinct from the candidate and taxonomy-validation figures.Source: §E.1 · p. 25
  • 784 of 1,403 papers (55.9%) belong to the five largest task fields; novel_view_synthesis has the highest coverage density at 6.5 data points per covered paper.Face-centric and generative_3d papers each exceed 6 data points per covered paper despite moderate corpus share.Source: §E.1 · p. 26, Figure 13 · p. 26, Figure 14 · p. 26
  • Reconstruction 3D concentrates 57.8% of claim mass in three Object Form leaves, while generative 2D places 30.2% on Semantic Match alone.Figure 15 displays the top 20 leaves; Figure 16 shows that generative_2d and scene_understanding are Prompt Match–dominated, while reconstruction_3d and gaussian_splatting are Object Form–dominated; novel_view_synthesis is most balanced across six axes.Source: Figure 15 · p. 27, Figure 16 · p. 27, §E.1 · p. 27
  • scene_understanding (11 active leaves) and optical_flow_stereo (8) are most specialised; novel_view_synthesis (28) and neural_radiance_field (33) are most diverse.Lower entropy means evaluation concentrates on fewer leaves, making omission of secondary criteria more consequential for selective comparison analysis.Source: §E.1 · pp. 27–28, Figure 17 · p. 28
  • human_pose_mesh papers concentrate on Correspondence and Anatomy, neural_radiance_field spans three leading leaves, and generative_3d favours Semantic Match and Shape.Across task types, the different top-five leaf profiles motivate field-signature analysis for selective comparison detection.Source: §E.1 · p. 28, Figure 18 · p. 29

E.2 Selective Comparison: Sampling Details

  • 200 of 789 eligible papers were sampled proportionally by task type, prioritising papers with more untested X-dimensions and comparison figures.The eligible pool averages 8.73 untested dimensions out of 10; the sample approximately preserves that distribution and spans all 20 task types.Source: §E.2 · p. 28
Loading 2610.00666v1…