PaperlayerView exact revision on arXiv

Source-linked AI summary

Can Frozen Hyperspherical Features Guide the Selection of Pseudo Masks?

Xinge Guo, Fengyang Xiao, Dingming Zhang, Yuhan Chen, Rihan Zhang, Xingjian Li, Tianyang Wang, Chunming He, Sina Farsiu

arXiv:2609.30080v1cs.CV

TL;DR

Foundation segmenters produce multiple plausible masks, but choosing among them may require another large model or an annotated quality head. SphereTrust ranks candidates using three properties of the partition they induce on frozen DINOv2 features, and also uses the same features in candidate-set student training. It outperforms the strongest evaluated external comparator on six of eight pools, while its training stage improves weighted F across the three MLLM anchor pools.

  • Problem

    Foundation segmenters return several plausible masks, while existing selection can require a second large model or a quality head fitted to annotated masks.

    Source: §1 · p. 1

  • Method

    SphereTrust scores candidate-induced partitions of normalized frozen DINOv2 patch features by angular contrast, foreground-mode coverage, and frame contact, then uses candidate scores, prototypes, and a cross-fitted second round in student training.

    Source: §1 · p. 2

  • Results

    On six of eight pools, SphereTrust exceeds the strongest evaluated external comparator by 1.7 to 9.3 percentage points in selected Dice; on two prompted camouflage pools, it is within 0.1 points of the candidate-derived DSS adaptation with a lower catastrophic-error rate.

    Source: §5 · p. 9

  • Takeaways & Limitations

    Across the tested ranking rules, candidate-set training with the prototype term and cross-fitted second round adds 2.6 to 10.7 points of weighted F over fixed-label training.

    Source: §1 · p. 2

    Selection needs two viable candidates per image, and its ViT patch-scale cues cannot resolve boundary detail within a patch.

    Source: §5 · p. 9

Abstract

from arXiv · show

Foundation segmenters such as SAM return several plausible masks for an unlabeled image, and a student trained on the wrong one inherits its errors. Choosing among them means querying a second large model or fitting a quality head to annotated masks. We show that a candidate can be judged by what it does to a frozen self-supervised backbone's features. Normalized DINOv2 patch features lie on a hypersphere, and a candidate mask splits that sphere in two. Based on this reading, we introduce SphereTrust, which scores each candidate by three properties of the split, the angular contrast between the two sides, the coverage of the foreground's appearance modes, and contact with the image frame, one for each of three common ways a mask fails, and ranks a pool in 0.55 s per image from the frozen features alone. On eight SAM and SAM3 candidate pools spanning camouflaged, salient, and dichotomous segmentation and camouflage under low light, SphereTrust exceeds the strongest evaluated external baseline on six pools by 1.7 to 9.3 percentage points in mean selected Dice. These comparisons include published selection rules and explicitly labeled adaptations of DSS and UCOD-MKD. On the two prompted camouflage pools, its mean selected Dice is within 0.1 percentage points of the candidate-derived DSS adaptation, with a lower catastrophic-error rate. Which cue carries the signal depends on the candidate pool. The same sphere also supports training. The leading candidates enter as a candidate set with their scores as priors, prototypes reorder them, and a cross-fitted second round completes the labels, raising weighted F by 4.5, 2.3, and 5.5 points over fixed-label training on the three MLLM anchor pools, with students competitive with published unsupervised methods on nineteen test sets.

1 INTRODUCTION

SphereTrust judges a candidate mask by how it partitions frozen DINOv2 features, rather than relying on a second large model or an annotated quality head. It also uses the candidate pool to support training, with strong selection and student-training results across the tested settings.

1 INTRODUCTION

  • Existing selectors use mask consensus, contours, or a second large model, while the generator’s own predicted IoU can rank masks poorly: .276 average Dice versus .685 for the best candidates on a camouflage pool.The comparison of best-candidate Dice and SAM predicted-IoU-selected Dice is reported in Appendix Tab. 7.Source: §1 · p. 1
  • A candidate mask partitions normalized frozen DINOv2 patch directions, and SphereTrust combines angular contrast, foreground-mode coverage, and frame contact to address wrong-object, omission, and leakage errors.The three terms are standardized within each image and summed with equal weights.Source: §1 · p. 2
  • On six of eight pools, SphereTrust exceeds the strongest evaluated external comparator by 1.7 to 9.3 percentage points in selected Dice; on two prompted camouflage pools, it is within 0.1 points of the candidate-derived DSS adaptation with lower catastrophic-error rate.It ranks candidates in 0.55 s per image; comparisons include published rules and explicitly labeled adaptations.Source: §1 · p. 2
  • The training stage retains leading distinct masks as a candidate set and uses their geometric scores as priors so the student can change targets as it learns.The three-stage training procedure also uses prototypes and a cross-fitted second round to complete labels.Source: §1 · p. 2
  • The study covers binary foreground/background DIS at ViT patch scale across camouflaged, salient, and dichotomous segmentation and synthetic low-light camouflage.The overview depicts geometric ranking followed by student training and a second round that adds a completed mask from a cross-fitted prediction.Source: §1 · p. 2, Figure 2 · p. 3

2 RELATED WORK

Prior work addresses pseudo-mask quality through confidence, consistency, consensus, contours, learned quality models, or related candidate-grading procedures. SphereTrust instead evaluates the partition induced by each candidate, while its training formulation connects to candidate-set supervision.

2 RELATED WORK

  • SAM exposes predicted IoU and SAMRefiner ranks candidates by it, while SAQ scores contours and learned quality models require labels or a reference set.Confidence can be miscalibrated, as noted by Guo et al. (2017).Source: §2 · p. 3
  • DSS scores a mask against a similarity map from a separately discovered region, whereas SphereTrust scores the bipartition induced by the candidate itself.The paper compares explicitly labeled adaptations on identical candidate pools.Source: §2 · p. 3
  • UCOD-MKD selects by generator confidence and grades quality using candidate agreement and, where applicable, structural checks; candidate-set learning instead resolves a set of labels by the model’s fit.The paper describes Eq. (5) as the spatial dense form, using generator candidates as the label set and the §3.1 score as prior.Source: §2 · p. 3
  • Unsupervised segmentation work also includes object discovery, pseudo-mask refinement, knowledge transfer, proposal agreement, and foundation-model priors, while mask-building frameworks are generators rather than selectors.The paper places published generator results beside its students rather than in the selector comparison.Source: §2 · p. 3

3 METHOD

SphereTrust ranks candidate masks through three cues computed from normalized frozen DINOv2 patch features. It then carries candidate scores into student training, where the student can revise its supervision and a cross-fitted second round adds a completed mask.

3.1 RANKING BY THE INDUCED PARTITION

  • Angular contrast F measures how far a mask partition separates the two mean feature directions; it is small when the mask splits one directional population and large when it separates two.The method uses normalized patch features on a 25 × 25 grid from 350 × 350 inputs.Source: §3.1 · p. 4, §3 · p. 4
  • Spherical coverage C averages recall across foreground appearance modes estimated from candidate-union features, using shared candidate patches and border-ring features to form foreground and background prototypes.An empty foreground mode scores zero; because coverage is nondecreasing as patches are added, it can reward leakage.Source: §3.1 · p. 4
  • Frame contact E is the fraction of the image’s outermost pixel ring occupied by the mask, and the rule subtracts this border term.The border is measured at the backbone’s 350 × 350 working resolution.Source: §3.1 · p. 4
  • The three terms are standardized within each image’s candidate set and summed with equal weights; masks that differ only within a patch score alike for the two patch-grid terms.The combined rule is untrained, and exact ties fall back to generator confidence.Source: §3.1 · p. 4
  • Figure 3 shows patch maps colored by nearest foreground appearance mode alongside three candidates and their F, C, E, and Dice values for one image per regime.The visual presents the three ranking cues on real candidates.Source: Figure 3 · p. 5

3.2 FROM GEOMETRIC RANKING TO CANDIDATE SET TRAINING

  • For training, leading distinct candidates are retained with geometric scores as priors, and the student balances each candidate’s score against its current BCE+Dice fit.After five fixed-label epochs in a 25-epoch run, the student may move to a lower-ranked candidate where its fit outweighs the score gap.Source: §3.2 · p. 5
  • A sphere term orders training candidates using cross-fitted foreground posteriors from von Mises–Fisher prototypes fitted on the other half of training images.The posterior’s mean inside a mask minus its mean outside forms the sphere term.Source: §3.2 · p. 5
  • In the second round, each image is predicted by a student trained on the other fold, candidates are rescored against that prediction, and a completed mask joins the four leading distinct candidates.The final student chooses per image between the completed mask and the plain candidates; the out-of-fold predictions share prototype information from the predicted fold.Source: §3.2 · p. 5

4 EXPERIMENTS

Experiments compare SphereTrust with other selectors on shared candidate pools, then test whether retaining and reordering candidates improves downstream student training. Selection is label-free, while downstream configuration uses annotated validation images.

4 EXPERIMENTS

  • The selector uses no ground-truth annotations, but downstream training configuration was selected using 1,085 annotated validation images.The evaluation uses shared candidate pools, and students share architecture, schedule, and three seeds.Source: §4 · p. 6

4.1 SELECTION ON A FIXED CANDIDATE POOL

  • On salient and dichotomous prompted pools, SphereTrust beats adapted DSS by 1.7 and 5.2 points; on prompted camouflage pools, it is within 0.1 points.The camouflage comparisons are .7873 against .7881 and .8241 against .8245; Table 1 covers the SAM3 and three MLLM-anchor pools.Source: §4.1 · p. 6, Table 1 · p. 6
  • On four unprompted SAM pools, SphereTrust exceeds the strongest evaluated external comparator by 2.5 to 9.3 points in selected Dice.Figure 5 compares SphereTrust with TokenCut and the candidate-derived DSS adaptation on pools where selectors must find the target among decoys.Source: §4.1 · p. 6, Figure 5 · p. 8
  • The full ranking score has the highest selected Dice among seven term combinations on all three MLLM anchor pools.It exceeds the best pair by .008 on COD, .036 on DIS, and .001 on SOD; coverage selects better than the area replacement on every prompted pool and the decoder pool.Source: §4.1 · pp. 6–7, Table 2 · p. 7

4.2 TRAINING THE STUDENT

  • The complete candidate-set training procedure raises weighted F by 4.5, 2.3, and 5.5 points over fixed-label training on COD, SOD, and DIS.Table 2 separates candidate ranking from candidate-set supervision and reports the ranking and training ablations.Source: §4.2 · p. 7, Table 2 · p. 7
  • Under matched training, the complete procedure improves all three tested scoring rules by 2.6 to 10.7 points over their own fixed-label baselines, with SphereTrust first in every regime and seed.Its mean advantage over the size control is 0.1 to 0.4 points, and matched score scaling removes that advantage in two regimes.Source: §4.2 · p. 9

5 DISCUSSION

SPHERETRUST improves candidate selection on six pools and nearly matches the candidate-derived DSS adaptation on two prompted camouflage pools. The discussion also identifies scope limits and cases where simpler or competing approaches perform well.

5 DISCUSSION

  • On six of eight pools, SPHERETRUST exceeds the strongest evaluated external comparator by 1.7 to 9.3 percentage points in selected Dice; on two prompted camouflage pools, it is within 0.1 points of candidate-derived DSS with a lower catastrophic-error rate.The comparisons include explicitly labeled adaptations of related methods.Source: §5 · p. 9
  • The candidate set, prototype term, and cross-fitted second round add 4.5, 2.3, and 5.5 points of weighted F under the authors’ ranking, and 2.6 to 10.7 points across three tested rankings.Source: §5 · p. 9
  • Selection requires two viable candidates per image, and patch-scale cues cannot resolve boundary detail within a patch, which dichotomous benchmarks score.Source: §5 · p. 9
  • On unprompted SAM output, area used instead of coverage selects as well or slightly better; on camouflage, DSS trains the better fixed-label student, while SPHERETRUST’s full procedure has a 0.18-point observed advantage.Source: §5 · p. 9
  • On COD10K, students trail the strongest published configurations; target identity was examined only on camouflage, and LL-COD is synthetic.Source: §5 · p. 10

REPRODUCIBILITY STATEMENT

The authors describe a planned public code release covering the experiments and report how configuration selection, final training, and testing were conducted.

REPRODUCIBILITY STATEMENT

  • The planned public GitHub release covers pool construction, candidate scoring and selection, data splits, prototype fitting, both training rounds, run configurations, candidate scores, and scripts producing tables and figures.Source: REPRODUCIBILITY STATEMENT · p. 10
  • Configuration was chosen using 1,085 annotated validation images held out of the selection process; reported students were then trained on complete pools, including those images, and scored once on test sets.Source: REPRODUCIBILITY STATEMENT · p. 10
  • The three final-student seeds share round-two candidate sets built from seed 0’s fold students, so their standard deviations describe final-stage variation conditional on those sets.Source: REPRODUCIBILITY STATEMENT · p. 10

AI USE STATEMENT

AI assistants supported writing, layout, routine scripting, and cross-checking, while the authors verified retained suggestions and take responsibility for the paper.

AI USE STATEMENT

  • AI assistants supported writing and language polishing, LaTeX compilation and layout, and elementary code assistance for routine data-loading, batch-execution, and result-table scripts.Source: AI USE STATEMENT · p. 10
  • Assistants also cross-checked reported numbers and literature comparisons, and authors verified feedback on comparison tables, iterative training, supervision ablation, and selection-result interpretation.Source: AI USE STATEMENT · p. 10
  • The authors produced the experiments and reported results, reviewed every retained suggestion against data and code, and take full responsibility for the paper.Source: AI USE STATEMENT · p. 10

APPENDICES

The appendices specify evaluation protocols, feature and candidate construction, training procedures, and anchor-box generation.

APPENDICES

  • Selection Dice uses native-resolution masks and ground truth binarized at 0.5 without resampling; the 350 × 350 grid is used only for features and patch-level terms.Source: §A.1 · p. 17
  • The selector uses frozen DINOv2 ViT-B/14 block-12 features on a 25 × 25 grid; SAM pools apply candidate filters, with prompted P2 images retaining 25 to 28 candidates.Source: §A.1 · p. 17
  • Foreground and border-ring prototypes and up to four fitted appearance modes feed equally weighted, within-image standardized score terms; student variants share a frozen backbone and training schedule.Source: §A.1 · p. 17
  • Round one trains with up to five distinct candidates scored by R + z(Γ); round two adds a completed mask and four leading distinct candidates, with fold artifacts shared across three seeds.Source: §A.1 · p. 18
  • Configuration was selected on fixed tenth-of-pool validation splits, while MLLM anchors use Qwen2.5-VL-7B-Instruct prompts for prominent and distinct salient objects, merged after duplicate removal.Source: §A.1 · p. 18

A.2 CANDIDATE POOLS, ELIGIBILITY, AND DATASETS

Candidate pools are fixed before selection, with benchmark-specific construction and eligibility filters; prompted regimes also use distinct anchor prompts and SAM3 phrase ladders.

A.2 CANDIDATE POOLS, ELIGIBILITY, AND DATASETS

  • P1 uses unprompted SAM candidates capped at 12, while P2 merges prompted candidates and supplies the training pools.P1 supports selection analysis because its candidates vary more in target identity and coverage; P2 is used for training.Source: §A.2 · p. 18
  • Selection includes only images with foreground and background annotations and at least two admissible candidates; eligibility ranges from 97.7% to 99.5% across four reported test sets.Downstream evaluation uses the full test sets, and the common candidate filters are applied identically to all selectors.Source: §A.2 · p. 18
  • The 15 evaluation sets comprise four camouflage sets, five salient-object sets, five DIS splits, and LL-COD, a synthetic low-light condition applied to COD10K test images.LL-COD uses fixed degradation settings and leaves the ground truth unchanged.Source: §A.2 · p. 19

A.3 PROMPTS AND PHRASE LADDERS

  • Camouflage prompts request up to three tight boxes around the likely camouflaged object, while salient and dichotomous prompts target foreground objects or regions.The prompt for non-camouflage regimes requests every distinct salient object; decoding is greedy, with box expansion and jitter used to form prompt boxes.Source: §A.3 · p. 19
  • SAM3 uses four phrases per regime, one at a time; its camouflage ladder supplies a pool, while salient and dichotomous ladders yielded weaker students than the MLLM anchor.The ladders differ by regime and retain each instance box with its phrase index.Source: §A.3 · p. 20

B.1 WHAT EACH TERM CONTRIBUTES

The analyses show that cue utility varies by pool and error type, while comparisons on fixed pools clarify both the unprompted SAM results and the limits of related scoring adaptations.

B.1 WHAT EACH TERM CONTRIBUTES

  • Removing angular contrast, the boundary prior, or spherical coverage reduces selected Dice by .126, .096, or .038, respectively, on the reported training subset.The boundary prior alone selects at .469 versus .629 for the full score; the F + E pair is .038 below the full score.Source: §B.1 · p. 20
  • Controlled corruptions show F responds most to wrong-object and leakage errors, C to omissions, and E little when errors do not reach the frame.Coverage responds negatively to leakage, so coverage alone can favor oversized masks.Source: §B.1 · p. 20, Figure 10 · p. 20
  • Area’s advantage changes with the pool: it wins on unprompted SAM output, whereas coverage wins on prompted SAM and another-generator pools.Coverage remains associated with Dice after controlling for F, E, and area; among similar-area candidates it selects the higher-Dice mask in 0.791–0.851 of pairs across three test sets.Source: §B.1 · pp. 20–21, Table 5 · p. 21

B.2 ANALYSIS OF THE SCORING RULE

  • On unprompted COD10K, normalized and raw-feature angular contrast both reach AUROC .946; the median-margin statistic reaches .934 on unit directions, .899 on raw features, and .752 using norms alone.Angular-contrast selected Dice rises from .162 to .575 across backbone blocks 3 to 12, and selection rises from .494 to .650 as candidate count increases from 3 to 12.Source: §B.2 · p. 21

C.1 STRESS TEST ON UNPROMPTED SAM OUTPUT

  • On four unprompted P1 test pools, SphereTrust beats the candidate-derived DSS adaptation by 2.5–9.3 Dice points and has the highest Top-1 and lowest catastrophic rate on each.The average candidate Dice is .0992–.1697, the oracle reaches .4856–.7878, and 91% of camouflage training images contain a decoy.Source: §C.1 · p. 21, Table 6 · p. 22

C.2 DETAILED SELECTOR AND STUDENT COMPARISONS

  • On the fixed P1 pool, structural checks raise generator-confidence selected Dice from .276 to .400, while the paired agreement statistic performs no better than chance (AUROC .459).On prompted pools, the same row reaches .715 to .889.Source: §C.2 · p. 22, Table 7 · p. 23

C.3 THE TWO CLOSEST SELECTORS ON A FIXED POOL

  • The candidate-derived DSS adaptation outperforms the anchor-box adaptation, while UCOD-MKD-inspired filtering is a distinct adaptation rather than the original procedure.Anchor-box scores are .620, .549, .839, and .606 versus .825, .788, .912, and .720 for candidate-derived maps; filtering changes merged-camouflage Dice from .728 to .772 and yields a 0.2-point gain on SAM3.Source: §C.3 · pp. 22–23
  • On prompted camouflage pools, SphereTrust and candidate-derived DSS are within 0.1 percentage points, with lower catastrophic-error rate for SphereTrust; SphereTrust leads by 1.7–9.3 points on other reported pools.The methods select different winners on 64% of the merged camouflage pool’s images.Source: §C.3 · p. 23

D.1 ABLATION OF THE TRAINING METHOD

The training ablations show that the full procedure outperforms simpler supervision variants, with gains depending on the segmentation regime. The tested configuration choices are relatively stable, while the reported seed variation is conditional on a shared upstream cache.

D.1 ABLATION OF THE TRAINING METHOD

  • The sphere term contributes on camouflage and dichotomous data and is neutral on salient data, although alone it reselects a better fixed label than R on every regime.Candidate-set training provides the camouflage gain; completion and cross fitting provide the salient and dichotomous gains.Source: §D.1 · p. 24
  • The full method exceeds selection-only and round-one training for every seed on every regime.The ablation compares the same architecture, images, and schedule while changing supervision.Source: Table 9 · p. 24, §D.1 · p. 24
  • The reported seed spreads describe final-student variation conditional on a shared seed-0 cache, not end-to-end pipeline variation.The fold students, out-of-fold maps, prototype fit, and round-two candidate sets were produced once and shared.Source: §D.1 · p. 24
  • Completion adds 1.7 and 1.6 points on salient and dichotomous data, but 0.2 on camouflage, where failures more often involve leaked background.Completion is a union, so it restores omitted parts but leaves leaked background in place.Source: §D.1 · p. 24
  • The oracle candidate is the pool’s best single mask, not an annotation ceiling, and the completed mask can extend beyond the candidate set.The completed mask joins the leading candidate with a confident region from a cross-fitted prediction.Source: §D.1 · p. 24

D.2 CHOICE OF THE TRAINING CONFIGURATION

  • Across test regimes, the four completion thresholds are within about one point; the two alternative orderings differ from the chosen one by at most 1.1 points.All four thresholds exceed round one alone on salient and dichotomous data and equal it on camouflage.Source: §D.2 · p. 24, Table 10 · p. 25

E DOWNSTREAM RESULTS BY TEST SET

Downstream results vary by task and metric: the student leads on several salient-object metrics, while other methods lead on others. Additional evaluations test fitted scoring rules and a separate candidate generator.

E DOWNSTREAM RESULTS BY TEST SET

  • The released Selfment checkpoint’s COD10K weighted F is .5507 under the stated protocol, versus .7540 as published; matching evaluators do not explain the gap.The authors report matching checkpoint and inference settings and agreement between evaluators on the same predictions.Source: E DOWNSTREAM RESULTS BY TEST SET · p. 25

F.1 FITTED SCORING RULES AND EVALUATION ACROSS DATASETS

  • Fitted scoring weights do not outperform equal weights: across evaluations, selected Dice is .647 versus .650 and .457 versus .478, while cross-half fitting ties at .617.The fitted predictor uses either three score-term features or eleven candidate statistics.Source: §F.1 · p. 25

E DOWNSTREAM RESULTS BY TEST SET

F.2 OTHER GENERATORS AND ANCHORS

  • On the DPT-lite candidate pool, the full score reaches .784 selected Dice, +.011 over the F + E pair at .773; generator confidence has better AURC, .126 versus .218.The pool averages 913 images from four sets outside the downstream suite.Source: §F.2 · pp. 25–26, Table 15 · p. 27

E DOWNSTREAM RESULTS BY TEST SET

F.3 ONE TRAINING PROCEDURE, THREE SELECTION RULES

A matched comparison separates candidate-selection rules from how their candidates are used for training. Under the complete procedure, SphereTrust leads both alternatives across the three regimes, although fixed-label winners differ by pool.

F.3 ONE TRAINING PROCEDURE, THREE SELECTION RULES

  • The three rules select different supervision on 44%–66% of images, and the comparison keeps each rule’s scoring, candidate order, and priors aligned.Images, student, schedule, hyperparameters, and seeds are matched across runs.Source: §F.3 · p. 27
  • Under the complete procedure, SphereTrust ranks first on every regime and seed, leading the size control by 0.1–0.4 points and adapted DSS by 0.2–1.0 points in regime means.The comparison matches student architecture, images, schedule, and seeds.Source: Table 16 · p. 28, §F.3 · p. 28
  • With fixed labels, adapted DSS trains the better camouflage student (.7624 versus .7526), while SphereTrust leads on salient objects (.8144 versus .7918) and dichotomous segmentation (.6047 versus .5464).The ordering depends on the pool even though the two leading rules select masks with similar mean camouflage quality.Source: §F.3 · p. 28
  • The complete procedure improves weighted F for every rule and narrows the spread among rules to 0.2, 1.0, and 0.7 points across camouflage, salient, and dichotomous regimes.The rule that starts furthest behind gains the most in each regime.Source: §F.3 · p. 28
  • After matching score scale, SphereTrust retains a 0.10–1.02-point margin over adapted DSS, but its comparison with the size control reverses on salient and dichotomous data.The authors therefore describe the size-control comparison as scale dependent and the two rules as level.Source: §F.3 · p. 28

G FOUR FURTHER BENCHMARKS

On four further prompted benchmarks, SphereTrust’s selection margins range from slightly negative to positive. The final students outperform selection-only students on all four sets, while the SAM3 and MLLM anchor comparison has a stated prompt-related caveat.

G FOUR FURTHER BENCHMARKS

  • The added benchmarks comprise PlantCamo for camouflage, SOD300 and MSRA-B for salient objects, and UHRSD for dichotomous segmentation.Overlapping SOD300 and MSRA-B photographs were removed.Source: G FOUR FURTHER BENCHMARKS · p. 29
  • The PlantCamo comparison does not isolate the effect of the animal-focused SAM3 phrase ladder from other differences between the training pools.PlantCamo targets plants, while the MLLM anchor prompt names no taxon.Source: G FOUR FURTHER BENCHMARKS · p. 29
  • Selected Dice margins over the strongest external comparator are −0.1, +1.8, +1.3, and +2.4 points on PlantCamo, SOD300, MSRA-B, and UHRSD, respectively.Candidate-derived DSS is the comparator on three sets and is 0.1 points ahead on PlantCamo, where SphereTrust leads on Top-1.Source: Table 18 · p. 29, G FOUR FURTHER BENCHMARKS · p. 29
  • Final students beat selection-only students on all four metrics across all four sets; F_w gains are +1.6, +5.5, +1.4, and +4.0 points.On PlantCamo, the SAM3-pool student trails the MLLM-anchor student by 9.3 points of F_w.Source: Table 19 · p. 29, G FOUR FURTHER BENCHMARKS · p. 29

Related papers

Loading 2609.30080v1…