CODA: Depth-Aligned Scene Completion and Object Decomposition from a Single RGB-D Image
Dongwon Son, Junhyek Han, Yoontae Cho, Minseok Lee, Hong-seok Choi, Jiwook Choi, Hyungjin Kim, Beomjoon Kim
cs.RO · cs.CV · cs.LG
Quick verdict
CODA repurposes TRELLIS.2 to complete a scene from one RGB-D view then split objects; strong ablations, 5.6 s runtime, but reports no significance tests.
Summary
CODA is a scene-first generative system that converts a single unsegmented RGB-D view into a complete 3D scene mesh, then splits that surface into the environment and individual movable objects, with no input instance masks. It fine-tunes frozen TRELLIS.2 via LoRA through three conditional flows — coarse 64³ occupancy, fine structured-latent geometry, and a texture flow repurposed to predict object-identity codes resolved by nearest-codebook assignment. Two mechanisms anchor generation to measurements: both geometry stages start from encoded observed-surface features (noise elsewhere), and depth-derived point maps back-project DINO features onto observed voxels before every DiT block. On HomebrewedDB and a custom 27-scene cluttered dataset, CODA reports complete-surface F@5 of 0.6789 and 0.6879 (about 14.5% above RecGen), pose retention of 0.785 and 0.766 under 2-second simulated gravity, and 5.59 seconds per scene on one RTX 4090 (roughly 17.9× faster than RecGen; ChatGPT Astra took 28.1 minutes). Ablations attribute large gains to grounding (F@5 0.1396 → 0.6789). Limitations: no significance tests or variance reporting; code release unmentioned (only the dataset is promised, after review); physical plausibility measured solely by gravity settling; and the diversity analysis uses oracle ground-truth sample selection in simulation only.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● low 2
5 strengths · 5 weaknesses · 5 author questions
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Completing one scene first and decomposing afterward outperforms object-first and scene-first baselines in complete-surface F@5 and gravity pose retention on HomebrewedDB and a custom cluttered dataset (0.6789/0.7852 and 0.6879/0.7661).
- Observed-surface initialization in both geometry stages plus depth-guided image-to-3D feature matching drive reconstruction accuracy, raising HomebrewedDB complete F@5 from 0.1396 (fine-tuned TRELLIS.2) to 0.6789.
- A pretrained texture flow repurposed to predict object-identity codes enables mask-free object separation; a learned-query decoder replacement collapses F@5 to 0.3748 and retention to 0.2010.
- Stochastic sampling yields diverse hidden-region completions while keeping observed surfaces consistent (CODA Dvis 0.068–0.173; Dhid rising 0.347 → 0.619 with occlusion).
- CODA reconstructs a scene in 5.59 s on a single RTX 4090, roughly 17.9× faster than RecGen and far faster than the 28.1-minute ChatGPT Astra workflow.
Strengths
Decomposition mechanism directly ablated
Replacing the repurposed texture flow with a PaSCo/Mask2Former-style learned-query decoder drops complete F@5 from 0.6789 to 0.3748 and pose retention from 0.7852 to 0.2010 (Sec. IV-C), isolating the contribution of identity-code generation.
Faithful grounding ablation ladder
Table II reconstructs prior art as ablations — Points-to-3D-style coarse-only init (0.3993), GenRecon-style matching (0.4501) — so each mechanism's marginal value is measurable up to the full model's 0.6789.
Segmentation independence verified against GT masks
Practical runtime with conservative baseline treatment
Occlusion-stratified stochastic analysis
Weaknesses
No statistical uncertainty anywhere
Reproducibility assets incomplete
Missing named baseline and baseline-tuning opacity
Physical validation is a proxy only
Under-specified decomposition details
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
CODA (arXiv 2609.25654; KAIST with Samsung Heavy Industries) targets single-view RGB-D reconstruction for robot manipulation: from one unsegmented RGB-D image with intrinsics (and an approximate camera pose constructed from a fitted support plane when Twc is absent), produce a complete scene mesh plus separate meshes for the environment and each movable object, without input instance masks or a known object count. The paper positions this “complete once, decompose afterward” ordering against object-first pipelines (RecGen, SAM 3D, ZeroGrasp, SceneComplete) that inherit segmentation errors, and scene-first methods (PaSCo, SIMstack, CoReNet) that output coarse voxels. Three hypotheses are stated (Sec. IV): H1 superiority in accuracy and gravity stability, H2 that depth grounding drives accuracy, H3 that stochastic generation diversifies hidden shapes while preserving observed geometry. Claims are quantified and mostly consistent with the tables, but no uncertainty quantification appears anywhere in the paper.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is CODA and what does it do?
CODA is a 3D generative model that takes a single RGB-D image of a cluttered scene (with camera calibration) and outputs a complete scene mesh plus separate meshes for the environment and each movable object — without needing any input segmentation masks. It is built on TRELLIS.2, fine-tuned with LoRA, and runs in about 5.6 seconds per scene on an RTX 4090.
How does CODA keep its generated 3D geometry consistent with the measured depth?
Through two mechanisms described in the paper: observed-surface initialization, which seeds both coarse and fine geometry generation with encoded features of the observed surface (random noise elsewhere, like 3D inpainting), and depth-guided feature matching, which back-projects DINO image features onto observed voxels via the depth point map and injects them before every transformer block. The paper's ablation shows these raise surface F@5 from 0.1396 to 0.6789 on HomebrewedDB.
Is CODA's code or dataset publicly available?
The paper states that the authors' custom evaluation dataset (27 cluttered scenes, 50 RGB-D views, 449 object-view targets) will be released after the double-anonymous review process. Code and trained model weights are not mentioned in the paper text; a project page exists at dongwon-son.github.io/coda-project-page.
How well does CODA perform compared to other methods?
On HomebrewedDB and a custom cluttered-scene dataset, CODA reports complete-surface F@5 of 0.6789 and 0.6879 (about 14.5% above RecGen), keeps 78.5% and 76.6% of evaluated objects in place during a 2-second simulated-gravity test, and is roughly 17.9x faster than RecGen. Note the paper reports no statistical significance tests for these comparisons.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures are present only as captions in the extracted text; qualitative evidence (Figs. 1, 5–7) could not be verified visually.
- Tables were transcribed from embedded text; individual cell values could contain extraction artifacts, though internal consistency checks passed (stated relative gains match table ratios, e.g. 0.6789/0.5929 ≈ +14.5%).
- Supplementary material, if any, was not included; details such as codebook size Q and per-dataset protocol nuances may live there.
- The project page and any released assets were not consulted (offline review).
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.