PaperlayerView exact revision on arXiv

Source-linked AI summary

LEGO-Anything: Coding Agents for 3D Scene Reconstruction

Xirui Li, Peng Shi, Mingwen Dong, Sheng Zhang, Zhuoyan Xu, Dongkyu Lee, Shuaichen Chang, Yi Xiang, Lin Pan, Jiarong Jiang

arXiv:2609.36380v1cs.CV

TL;DR

Single-image 3D reconstruction is often represented as a rendering or fixed 3D output, limiting direct inspection, editing, and querying of the result. LEGO-Anything builds executable Blender scene programs iteratively and evaluates them with LEGO-Bench, LEGO-Plugin, and LEGO-World; results show strong benchmark performance by GPT-6-astra but persistent fidelity gaps and limited vision-task precision.

  • Problem

    Existing reconstruction approaches rarely produce an explicit scene program that can be checked against the input and revised, limiting direct inspection, editing, and querying of the reconstruction.

    Source: §1 · p. 1

  • Method

    LEGO-Anything has a coding agent iteratively write and execute Blender programs, inspect scenes and renderings, and revise them, with benchmark, trajectory, and vision-task evaluations.

    Source: §1 · pp. 1–2

  • Results

    53.4% indoor and 39.6% outdoor scores make GPT-6-astra the strongest overall evaluated agent, while scene readouts show non-trivial performance across detection, segmentation, and depth but remain below specialized vision models.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Executable scene programs are a promising, measurable, and diagnosable representation of natural images, but current coding-agent reconstructions are not yet sufficiently precise.

    Source: §7 · p. 10

    Current agents reliably deliver valid scene artifacts but remain limited in geometric fidelity, and frozen-scene vision readouts fall well below specialized models.

    Source: §4.3 · p. 6, §7 · p. 10

Abstract

from arXiv · show

A 3D scene reconstructed from a single image is most useful when represented not as a rendering or a fixed 3D output, but as an explicit scene program whose execution yields a scene that can be inspected, edited, and queried. We present LEGO-Anything, an Image-to-Code framework in which a coding agent iteratively writes and executes Blender code, inspects scenes and renderings, and revises the program. To evaluate end-to-end scene recovery, we introduce LEGO-Bench, a simulator-grounded benchmark with 208 images from 104 diverse indoor and outdoor scenes. LEGO-Bench separately scores artifact validity, visible-surface geometry, and rendered appearance. Its simulator-grounded design enables extensibility and precise automatic evaluation. Among evaluated agents, GPT-6-astra achieves the strongest overall results, with 53.4% indoor and 39.6% outdoor scores, yet substantial gaps remain between delivering valid scene artifacts and faithfully recovering scene geometry and appearance. Analysis of agent construction trajectories reveals three recurring issues: weak scene initialization, regressive edits during iteration, and unreliable self-evaluation. These findings motivate LEGO-Plugin, a training-free harness plugin for more controlled iterative scene construction, which improves all six evaluated models, with relative gains of up to 62.7% in overall score. Finally, we test whether reconstructed scenes can represent natural images and support vision tasks. In LEGO-World, we derive object detections, instance masks, and relative depth as deterministic queries on scenes reconstructed by GPT-6-astra. These readouts show non-trivial performance across all three tasks but fall well short of specialized vision models, suggesting that program-constructed scenes from current coding agents are a promising but not yet sufficiently precise representation of natural images.

1 Introduction

LEGO-Anything treats single-image 3D reconstruction as iterative construction of an explicit, executable scene program. The paper introduces a benchmark for measuring reconstruction, diagnoses agent failures and tests a harness plugin, then evaluates scene programs as representations for vision tasks.

Motivation and framework

  • LEGO-Anything represents reconstruction as an executable scene program whose explicit objects, geometry, layout, and camera can be run, inspected, edited, and queried.Unlike modular pipelines or fixed 3D outputs, the iterative formulation executes each step, compares it with the input image, and refines it.Source: §1 · p. 1

Benchmark and evaluation

  • LEGO-Bench evaluates 208 images from 104 diverse indoor and outdoor scenes, separately scoring artifact validity, visible-surface geometry, and rendered appearance.Its simulator-grounded design retains private geometry, depth, and instance masks for precise automatic evaluation.Source: §1 · p. 2, Table 1 · p. 3
  • 53.4% indoor and 39.6% outdoor scores make GPT-6-astra the strongest overall evaluated agent, while substantial gaps remain between artifact validity and faithful geometry and appearance.LEGO-Bench distinguishes failure to deliver a valid artifact from failures of geometric and visual fidelity.Source: §1 · p. 2

Failure diagnosis and LEGO-Plugin

  • Trajectory analysis identifies weak scene initialization, regressive edits, and unreliable self-evaluation as recurring failures that motivate LEGO-Plugin.Its three modules target those failures through reference-grounded initialization, evidence-grounded refinement, and protection of previously correct progress.Source: §1 · p. 2
  • LEGO-Plugin improves all six evaluated models on the LEGO-Bench Office subset, with relative gains of up to 62.7%.It plugs into existing harnesses as tools, skills, and runtime hooks rather than introducing a separate planner.Source: §1 · p. 2

Scene programs for vision

  • LEGO-World derives detections, instance masks, and relative depth from reconstructed scenes, with non-trivial performance that remains well below state-of-the-art vision models.The results suggest that current coding-agent scene programs are promising but not yet sufficiently precise representations of natural images.Source: §1 · p. 2

2 Related Work

Related work covers coding agents and recent visual and 3D program-based systems, then reviews scene-construction methods and benchmarks under differing assumptions. The paper positions its focus as single-image, end-to-end executable scene reconstruction and evaluation.

2.1 Coding Agents

  • Recent visual and 3D coding-agent systems include VIGA, SEIG, ArtiCraft, and Code-CoT, while this work studies general-purpose agents for single-image executable scene construction.Its emphasis is scene fidelity, iterative failure modes, and the perceptual sufficiency of frozen reconstructed scenes.Source: §2.1 · p. 3

2.2 3D Scene Construction and Evaluation

  • Procedural environment generation, image-conditioned recovery, editable RGB-D scenes, and simulation-ready articulated assets address components of scene construction but do not directly test faithful executable reconstruction from one natural image.The cited directions include procedural systems (Deitke et al., 2022; Raistrick et al., 2023, 2024; Deng et al., 2024) and image-conditioned methods (Cao et al., 2026; Team et al., 2026).Source: §2.2 · p. 3
  • LEGO-Bench targets single-image, end-to-end executable scene reconstruction across diverse environments and separates artifact validity, visible-surface reconstruction, and rendered appearance.Prior benchmarks cover procedural objects, parametric parts, text-specified worlds, or calibrated multi-view indoor reconstruction rather than this full setting.Source: §2.2 · p. 3

3 LEGO-Anything: Scene Reconstruction as Image-to-Code

LEGO-Anything formulates single-image reconstruction as an Image-to-Code task: a general-purpose coding agent uses Blender to construct an executable scene program. Rather than producing code in one shot, it iterates through editing, execution, and inspection before submitting an artifact for evaluation and downstream queries.

Scene reconstruction as Image-to-Code

  • Given a single RGB image, a general-purpose coding agent interacts with Blender to produce a scene program whose execution constructs a 3D scene.The program, rather than a fixed 3D output or rendering alone, is the central reconstruction artifact.Source: §3 · p. 3, Figure 2 · p. 4

Iterative construction

  • The agent alternates between editing code, executing it, and inspecting the resulting scene and renderings, producing a trajectory of intermediate programs, scenes, and observations.The final submission includes the scene, its export, and a rendered view.Source: §3 · p. 4

Evaluation and downstream use

  • The executable artifact is evaluated for fidelity and can be queried for downstream perception, while the workflow proceeds from reference interpretation to validated final artifacts.The paper evaluates final scenes in LEGO-Bench, analyzes trajectories to design LEGO-Plugin, and queries frozen scenes in LEGO-World.Source: §3 · p. 4, Figure 2 · p. 4

4 LEGO-Bench: A Simulator-Grounded Benchmark for End-to-End Scene Reconstruction

LEGO-Bench evaluates single-image, end-to-end scene reconstruction using natural-image-style views rendered from simulator scenes with retained ground truth. It separates artifact validity, visible-surface geometry, and appearance, and reveals that strong artifact delivery does not ensure faithful reconstruction.

4.1 Benchmark Design and Construction

  • Human review and collision checks gate candidate scenes, while the asset-and-scene-specification pipeline supports adding environments, difficulty levels, and views without changing evaluation.Figure 3 depicts Fab asset mining and assembly, followed by human verification that triggers remaking or annotation before acceptance.Source: §4.1 · p. 5, Figure 3 · p. 5
  • 208 RGB inputs from 104 scenes span 8 environments and 17 themes, with fixed scene conditions across nested difficulty tiers as visible content increases.The benchmark renders professionally authored simulator scenes to retain private geometry, object identities, camera parameters, depth, and instance masks for evaluation.Source: §4.1 · pp. 4–5

4.2 Evaluation Metrics

  • Validity checks artifact usability, Reconstruction measures visible-surface geometry, and Appearance measures rendered-image similarity; invalid submissions receive zero overall score.Valid submissions receive the mean of Reconstruction and Appearance, while the benchmark averages over all attempted cases, including failures.Source: §4.2 · pp. 5–6

4.3 Main Results

  • 53.4%/39.6% overall scores make GPT-6-astra the strongest evaluated configuration indoors/outdoors, while Validity is near-saturated and fidelity varies sharply.Three of six baselines lack outdoor support, while coding agents run on both splits; fidelity drops outdoors for every model.Source: Table 3 · p. 6, §4.3 · p. 6
  • 99.5% Validity persists across complexity tiers, while overall score falls from 24.6% on Easy to 20.5% on Hard and outdoor Reconstruction drops from 18.4% to 12.7%.Reconstruction and Appearance decline consistently on both splits as complexity increases.Source: Table 4 · p. 7, §4.3 · p. 7
  • 32.3% to 61.8% for GPT-6-astra, 21.3% to 39.7% for GPT-6-sol, and 14.4% to 21.2% for GPT-6-luna overall scores accompany increased reasoning effort.GPT-5.6 models show weak or non-monotonic changes; astra approaches saturation toward XHigh while sol continues improving.Source: Figure 4 · p. 7, §4.3 · p. 7

5 LEGO-Plugin: From Failure Diagnosis to Reliable Scene Reconstruction

Trajectory analysis identifies weak initialization, regressive edits, and unreliable self-evaluation as recurring construction failures. LEGO-Plugin targets these failures with three training-free modules and improves every evaluated model on the Office subset.

5.1 Diagnosing Construction Failures

  • GPT-6-astra reaches an evaluable scene within roughly the first tenth of its budget, while GPT-5.6-sol needs about one fifth; 29.6% of GPT-5.6-sol edits lower its score.Its final scene trails its best intermediate scene by 3.2 points, with later edits damaging geometry, camera settings, or stronger prior scenes.Source: §5.1 · pp. 7–8
  • Self-judgments agree with metric direction on 45.8% of Reconstruction and 62.2% of Appearance comparisons, and offer no consistent advantage over cross-model judgments.Judges are near or below chance on geometry, motivating refinement grounded in deterministic evidence rather than self-assessment.Source: §5.1 · p. 8, Figure 6 · p. 8

5.2 LEGO-Plugin

  • LEGO-Plugin is a training-free control layer exposed through tools, workflow skills, and runtime hooks, rather than a separate planner or memory module.Enhanced Initialization derives a reference-aligned starting scene; Grounded Refinement uses reference object regions, relative depth, and explicit residuals; Version Control accepts, repairs, or rolls back candidate updates.Source: §5.2 · pp. 8–9

5.3 LEGO-Plugin Evaluation

  • LEGO-Plugin improves all six models on the Office subset, with relative gains of 55–63% for GPT-5.6 models and 2.1% for GPT-6-astra.Gains are inversely related to base performance; the plugin narrows but does not close the gap to the strongest model.Source: Figure 8 · p. 9, §5.3 · p. 9

6 LEGO-World: Reconstructed Scenes as Visual Representations

LEGO-World tests whether one reconstructed scene can serve as a source for standard vision-task readouts. The readouts perform non-trivially across detection, segmentation, and depth, but remain substantially behind specialist models.

LEGO-World Evaluation

  • Frozen reconstructed scenes yield 30.1 box AP, 14.8 mask AP, and 0.155 AbsRel, compared with 59.9 for DINO, 54.0 for SAM 3, and 0.078 for Depth Anything 3, respectively.The evaluation uses GPT-6-astra scenes without LEGO-Plugin and 100 sampled images per dataset track; box and mask AP use equal confidence because scene exports lack calibrated confidences.Source: §6 · p. 10, Table 5 · p. 10
  • From each frozen scene, visible instances are projected into 2D boxes and masks, while camera-space depth renderings provide relative depth.The scenes are reconstructed from natural images by GPT-6-astra, and the agent labels objects using each target dataset’s category set.Source: §6 · p. 10
  • Scene readouts support detection, segmentation, and depth without task-specific training, but remain well below specialist models and are not yet sufficiently precise representations of natural images.The readouts are deterministic queries on the same frozen scene.Source: §6 · p. 10

7 Conclusion

The conclusion presents executable scene programs as a promising, measurable, and diagnosable target for coding-agent reconstruction, while noting that fidelity and downstream readout precision remain limited. The related-work discussion situates this task among procedural systems, object and scene reconstruction, coding agents, and executable-3D benchmarks.

7 Conclusion

  • LEGO-Anything reconstructs single-image scenes as editable, executable programs; LEGO-Bench separates validity, visible-surface geometry, and appearance, while LEGO-World queries frozen scenes for vision tasks.The conclusion reports that agents reliably deliver valid artifacts but remain limited in geometric fidelity, and that scene readouts remain well below specialist models.Source: §7 · p. 10

A.1 Procedural Scene Generation and Simulation Environments

  • Procedural systems generate environments from structured specifications and reusable assets; LEGO-Anything instead tests whether an agent can infer such a specification from one reference image.Examples include ProcTHOR, Infinigen, Infinigen Indoors, and CityCraft.Source: §A.1 · p. 17

A.2 Object Reconstruction and Asset Generation

  • Object-level reconstruction strengthens geometry and appearance priors, while articulated-asset work recovers movable parts or simulation-ready structure; LEGO-Anything evaluates assembly of full executable scenes.The cited object-level methods do not by themselves determine how objects should be selected, positioned, articulated, and integrated into a coherent scene.Source: §A.2 · p. 17

A.3 Scene Reconstruction and Editable 3D Intermediates

  • Holistic reconstruction methods recover multiple objects, geometry, semantics, or layout from partial evidence, while BlenderFusion uses explicit scene state for visual editing and compositing.These lines are adjacent to LEGO-Anything’s focus on executable, inspectable scenes rather than fixed scene representations.Source: §A.3 · p. 17

A.4 Coding Agents and Harnesses

  • Coding-agent work uses code for planning, tool interaction, and execution feedback; LEGO-Anything differs by evaluating end-to-end image-conditioned reconstruction under hidden scene ground truth.The related work includes harness systems and visual or 3D coding agents, including LL3M’s Blender-script refinement.Source: §A.4 · p. 18

A.5 3D Construction Benchmarks

This block situates LEGO-Bench among executable-3D benchmarks and reports supplementary analyses of cost, construction trajectories, and overhead-view reconstruction. The supplementary results highlight quality–cost trade-offs, non-monotonic trajectories, and persistent geometry gaps.

A.5 3D Construction Benchmarks

  • LEGO-Bench evaluates image-conditioned, scene-level, end-to-end inference of a complete executable scene from visual evidence, unlike benchmarks that generate code from textual or parametric specifications.Buildee emphasizes interactive exploration and simulator tasks, whereas LEGO-Bench fixes a single RGB observation and evaluates final scene artifacts across diverse scene families.Source: §A.5 · p. 18

B.1 Cost–Performance Pareto Frontier

  • GPT-5.6-luna is the cheapest configuration, while GPT-6-astra reaches the highest benchmark score at substantially higher spend; GPT-5.6-sol and GPT-5.6-terra lie off the quality–cost frontier.GPT-6-luna improves sharply over the lowest-cost regime, and GPT-6-sol adds a quality tier at moderate cost.Source: §B.1 · p. 18, Figure 10 · p. 19

B.2 Construction Trajectory Diagnostics

  • GPT-6-astra reaches its first evaluable scene later in absolute time but earlier relative to its runtime, and has the lowest regression rate and best-to-final regret.All four models remain non-monotonic; even the strongest can regress sharply late in construction.Source: §B.2 · p. 19, Figure 11 · p. 20

B.3 Bird’s-Eye Reconstruction Stress Test

  • GPT-6-astra is strongest on the Bird’s-Eye stress test, and the GPT-6 family substantially outperforms GPT-5.6; all models retain a gap between artifact delivery and strict-geometry reconstruction.The unpaired overhead-view cases are reported as a stricter stress test, separate from the paired scene-complexity analysis.Source: §B.3 · pp. 19–20

B.4 Pilot Scene-Level Articulation Evaluation on LEGO-Bench

The pilot applies a canonical articulation manifest to a frozen Office scene and still scores zero; the surrounding benchmark materials describe scene capture, controlled complexity, and representative agent outputs.

B.4 Pilot Scene-Level Articulation Evaluation on LEGO-Bench

  • The pilot classifies each joint requirement as missing object, part, or joint, wrong joint type, or correct, then macro-averages object scores and assigns zero for false motion on confirmed-static objects.The per-object score weights correct requirements twice; scene scores also include predictions assigning movable joints to matched static objects.Source: §B.4 · p. 21
  • With all 14 articulated targets scoreable and 12 static controls retained, the frozen Office submission still scores Art = 0, reflecting its prediction rather than missing ground truth.The earlier partial adapter left 12 articulated targets unavailable.Source: §B.4 · p. 21

C Demos on LEGO-Bench

  • Figure 13 pairs each reference image with evaluator-rendered final artifacts from six coding-agent configurations on the same evaluation split.The examples include the Havana rum shop street scene.Source: C Demos on LEGO-Bench · p. 22

D.1 Scene Construction and Capture

  • LychSim constructs and captures benchmark scenes, while collision and stability checks and human review screen candidate compositions; task inputs expose images and camera metadata but keep evaluation ground truth private.The capture backend records annotations used by the evaluator.Source: §D.1 · p. 22

D.2 Controlled Scene Complexity

  • Difficulty is controlled through nested object additions within matched scene families, holding shared scene and object properties fixed; the unmatched Aerial XHard split is excluded from paired complexity analysis.Retained objects must preserve support consistency and appear in at least one reference view.Source: §D.2 · p. 23

D.3 Dataset Statistics

The section documents the benchmark’s asset statistics and annotation process, then specifies how validity, visible geometry, appearance, and failures are evaluated and aggregated.

D.3 Dataset Statistics

  • Object scales span small manipulable items through architectural and city-scale structures, while Table 7 covers canonical articulation requirements parsed from human-accepted URDFs.The figure reports object-size distributions; the table caption identifies its manifest coverage scope.Source: Figure 16 · p. 24, Table 7 · p. 24

D.4 Articulated-Asset Annotation

  • After 15 human annotation rounds, all 443 canonical assets were resolved: 368 static and 75 articulated, with accepted annotations linked to reviewed revisions and motion-validation evidence.The workflow uses automated proposals and deterministic checks, but every output requires human review and execution-verified motion before entering evaluation.Source: §D.4 · p. 24

E Metric Implementation and Validation

E.1 Artifact Checks and Reported Validity

  • Artifact validity requires reloadable non-empty Blender and GLB scenes plus a decodable, non-degenerate render; unresolved headline-evaluation failures count as invalid in the main table.Artifact-invalid attempts remain in the denominator and receive zero validity-gated Final score.Source: §E.1 · p. 25

E.2 Reconstruction Alignment and Scoring

  • Headline Reconstruction compares visible surfaces per object without registration, using a depth-relative 5% tolerance and equal-weight macro-averaging; missing surfaces lower recall and extra in-mask geometry lowers precision.The evaluator uses the active camera frame and caps the whole-scene representation at 100,000 points.Source: §E.2 · pp. 25–27

E.3 Evaluator-Rendered Appearance

  • Appearance is the percentage of pixels within a 30-level channel difference on a fresh evaluator render, making it sensitive to lighting, materials, camera, and geometry rather than a perceptual or independent geometry measure.On a fixed six-task subset, GPT-6-astra stayed ahead of GPT-5.6-sol across tested thresholds, appearance tolerances, and point budgets.Source: §E.3 · p. 27

E.4 Aggregation and Failure Accounting

  • The overall score averages validity-gated Reconstruction and Appearance over all attempted cases, assigning zero to artifact-invalid submissions and unresolved evaluator failures.Source: §E.4 · p. 28

E.5 Human-Metric Alignment

A blinded human study finds substantial annotator agreement and reports close compatibility between human preferences and the headline metrics; the section also describes articulation scoring and experiment setup.

E.5 Human-Metric Alignment

  • Human-metric compatibility is 83.7% when near-tied metric scores and balanced human votes count as Both OK; annotator compatibility agreement is 86.6% overall, 88.9% for Reconstruction, and 84.3% for Appearance.The study reports six annotators, 240 samples, and 960 local areas.Source: §E.5 · p. 28

E.6 Articulation Protocol

  • Articulation scoring combines movable-part structure F1, matched-part surface quality, and quality-weighted joint F1; assigning motion to a human-confirmed static object is an explicit error.Movable parts are matched in an object-centric frame because global scene pose is not the functional target.Source: §E.6 · pp. 28–29

F.1 Common Runtime and Artifacts

  • Main coding-agent campaigns use isolated tasks, a single public RGB image, and three independent runs per configuration, with failed or incomplete artifacts remaining in evaluation.The agent receives no private depth, segmentation, or metric feedback.Source: §F.1 · p. 29

F.2 Method Configurations and Information Access

  • The main table compares end-to-end agents and baselines using the same frozen evaluator, but inference budgets and information access are not equalized; baseline task support varies.Some systems are evaluated only indoors, while others cover both Indoor and Outdoor splits.Source: §F.2 · p. 29

F.3 VLM Judge Setup and Reference Results

The section reports VLM judge agreement and describes the reasoning-effort sweep and the bounded plugin services for scene initialization, editing, validation, and execution records.

F.3 VLM Judge Setup and Reference Results

  • Across 2,160 primary judgments, there were no terminal failures; order-swap consistency ranged from 69.4% to 100.0% for Reconstruction and 61.1% to 83.3% for Appearance.Six models served as builders and judges; the judge saw only reference images and candidate renders.Source: §F.3 · pp. 29–30

F.4 Reasoning-Effort Protocol

  • The reasoning-effort sweep evaluates six models at four settings on 42 Office tasks; elapsed times are not a controlled measure of reasoning cost because concurrency and shared-host disk contention varied.Infrastructure failures trigger reruns of the same configuration, not best-of selection.Source: §F.4 · p. 30

G.1 Runtime Boundary

  • LEGO-Plugin leaves planning to the agent and general editing to Blender MCP, exposing only bounded measurement, validation, transaction, and evidence operations with no unrestricted code execution.The plugin changes scenes only after an explicit agent call.Source: §G.1 · p. 31

G.2 Shared Reference Evidence

  • Initialization and refinement use reference-image evidence from VGGT, with optional SAM 3 masks and Depth Anything V2 relative depth; the agent decides which Blender entity each region represents.Source: §G.2 · p. 31

G.3 Enhanced Initialization

  • Enhanced Initialization estimates a gravity-aligned Z-up room frame and camera, then produces a per-object plan of normalized position, projected extent, relative depth, and observability without estimating metric depth or absolute size.Source: §G.3 · p. 31

G.4 Version Control

  • Version Control snapshots scene state before edits and commits only allowed, non-empty changes that pass hard requirements; otherwise it rolls back, with at most eight transactions and two rollbacks per case.Source: §G.4 · p. 31

G.5 Grounded Refinement

  • Grounded Refinement checks explicit object or collection requirements, blocks acceptance on hard failures or missing evidence, and allows at most two correction rounds before commit or rollback.Any later scene change invalidates the validation report tied to the earlier scene state.Source: §G.5 · p. 31

G.6 Execution Records

  • Execution records log each plugin call and scene signature; runtime hooks block finishing with stale validations, unresolved transactions, or unreviewed renders, then export scene artifacts and a render.Source: §G.6 · p. 31

H.1 Task Inputs and Frozen-Scene Readouts

The evaluation freezes reconstructed scene artifacts before task readout and reports detection, segmentation, and relative-depth comparisons with coverage and runtime choices made explicit.

H.1 Task Inputs and Frozen-Scene Readouts

  • Each scene is constructed before evaluation and frozen for COCO detection, LVIS instance segmentation, and ETH3D relative depth; the tracks do not evaluate all three annotation types on the same image.Readouts project scene geometry into visible-object boxes and masks and render camera-space depth; ground-truth annotations and metric feedback remain verifier-only.Source: §H.1 · p. 32

H.2 Detection and Segmentation: Confidence and Missing Predictions

  • Scene-based box and mask AP assigns every prediction confidence 1.0 with deterministic tie ordering, making it a fixed-order compatibility diagnostic rather than calibrated ranking.Direct baselines are DINO-R50-4scale-24ep for boxes and SAM 3 for masks; both scene-based AP tracks score 94 of 100 attempted images while official subset AP retains all 100.Source: §H.2 · p. 32

H.3 Relative Depth and Coverage

  • 0.07845 versus 0.15540 AbsRel on the same 99 valid images for archived direct and scene-based predictions, respectively.The comparison uses relative-depth scale-and-shift alignment, not absolute metric scale.Source: §H.3 · p. 32

I Evaluation Infrastructure and Reproducibility

  • Harbor supports task expansion, repetitions, logging, artifact collection, verifier execution, and aggregation, while formal runs use a common isolated Docker/Compose runtime and artifact policy.Reported runs are identified by frozen manifests; Codex uses the Bedrock provider in final campaigns, and the Hermes wrapper supports Bedrock Converse with session export.Source: I Evaluation Infrastructure and… · p. 33
Loading 2609.36380v1…