{"id":"e615be23-87be-4f09-99c1-0368f3295355","arxiv_id":"2607.11118","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Geometry-guided hallucination of opaque textures from transparent regions lets off-the-shelf depth and reconstruction models recover accurate surfaces without retraining.","lead":"GHOST is a preprocessing pipeline that paints transparent objects as opaque, textured surfaces so ordinary depth and 3D models work on glass without retraining. It matters because robots and reconstruction systems still fail on everyday transparent materials, and this approach reuses existing SOTA models instead of building specialized ones.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"No significant objection identified beyond the reader's already-flagged weak-refraction assumption.","rationale":"The reader correctly isolates the weak-refraction assumption as the softest modeling choice and still assigns CONDITIONAL with high confidence. My re-examination of the pipeline, loss formulations (Eqs. 7–18), and experimental design finds no stronger or additional load-bearing flaw: the ablations demonstrate necessity of every stage, the gains appear consistently across five independent foundation models, and the method is presented as a practical preprocessor rather than a universal physical solver. The concrete test above simply operationalizes the same limitation the reader already noted. Therefore the verdict remains CONDITIONAL; no adjustment is warranted.","tokens_in":16186,"tokens_out":433,"duration_ms":6533,"concrete_test":"On a held-out subset of thick or high-index glass objects (or synthetic renders with known refractive index ≫1.5), recompute TransDecomp alpha/foreground and the full GHOST pipeline; if the resulting depth/normal metrics in Table 2 degrade to or below the raw-image baseline, the weak-refraction premise fails for that regime and the claim must be scoped to thin-walled objects.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (GHOST-preprocessed RGB restores photometric cues so that off-the-shelf depth/reconstruction models succeed on transparent objects without retraining) is supported by the multi-model tables (2–4), ablations (5–6) that correctly show partial pipelines degrade performance, and zero-shot TROS results. The only modeling premise that could undermine the claim is the weak-refraction alpha-blending approximation used to train TransDecomp (Sec. 3.3, Snell’s-law argument n1≈n2 and Eq. 7). The paper itself notes that strong refraction yields lower alpha and that the pipeline then falls back to surface-texture synthesis; this is a stated limitation rather than a hidden inconsistency. No other internal contradiction or experimental confound rises to load-bearing status.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"GHOST is a four-stage preprocessing pipeline (TransDINO segmentation, TransDecomp alpha/foreground decomposition, DAF-Net surface-normal estimation, GeoSemTransNet geometry-conditioned texture synthesis) that converts transparent image regions into opaque, structure-preserving RGB so that frozen off-the-shelf depth estimators and feed-forward reconstructors can operate without retraining. The central empirical claim is that the resulting images restore photometric cues sufficiently for large, consistent gains on ClearGrasp/ClearPose (Tables 2–4) and zero-shot transfer to TROS (Table 7), with ablations (Tables 5–6) showing that partial pipelines degrade performance.","tokens_in":16452,"tokens_out":919,"duration_ms":11665,"significance":"If the reported gains hold under broader conditions, the work supplies a practical, model-agnostic remedy for a long-standing failure mode of Lambertian 3D vision systems. The multi-model evaluation (DA3, MoGe2, DepthPro, Metric3Dv2, VGGT, DUSt3R, π3, MapAnything) and the demonstration that generic models plus GHOST can match or exceed specialized transparent-object estimators (Table 3) are concrete strengths. The modular design and explicit ablations make the contribution reusable and falsifiable. The principal modeling premise—the weak-refraction alpha-blending approximation—is stated rather than hidden, so the paper remains useful even where that premise is only partially valid.","major_comments":[{"comment":"Section 3.3 (Snell’s-law argument n1≈n2 and Eq. 7) trains TransDecomp under a weak-refraction alpha-blending model whose background prior B is produced offline by InpaintAnything. The paper notes that strong refraction yields lower alpha and that GeoSemTransNet then falls back to surface-texture synthesis, but no quantitative stratification by refractive strength, wall thickness, or multi-layer stacking is reported. Because every subsequent stage depends on the quality of α and Fg, a controlled failure-mode analysis (or an explicit statement of the operating envelope) is needed to bound the central claim that the synthesized opaque images remain geometrically faithful.","section":null},{"comment":"Tables 2–4 and 7 report large metric improvements, yet the evaluation is confined to ClearGrasp/ClearPose (and a TROS zero-shot set) with images resized to 256×256. No multi-object occlusion, extreme lighting, or outdoor transparent-surface experiments are shown. A modest expansion of the test distribution—or an explicit limitation paragraph quantifying domain shift—would strengthen that the gains are not dataset-specific.","section":null}],"minor_comments":[{"comment":"Figure 1 caption and the surrounding text use both “GeoSemTransNet” and “GeoSem-TransNet”; a single spelling should be fixed throughout.","section":null},{"comment":"Equation (1) and the subsequent fusion equations would benefit from an explicit statement of the feature dimensions after the DINOv3 projection, so that the cross-attention implementation is reproducible.","section":null},{"comment":"Table 1 reports mIoU on four datasets, but the training protocol for the multi-view ClearPose subset (every 500th frame) is only briefly mentioned; a short note on whether temporal leakage was checked would improve clarity.","section":null},{"comment":"The loss-weight choices (λsfs=10, λgrad=5, etc.) are given empirically; a one-sentence sensitivity remark or a pointer to the appendix would help readers who wish to re-implement.","section":null},{"comment":"Several recent transparent-object depth/normal works (e.g., DKT) appear only in the comparison tables; a short related-work paragraph situating GHOST relative to diffusion-based alternatives would complete the literature framing.","section":null}],"recommendation":"minor_revision","confidential_remarks":"The manuscript is a solid systems paper whose empirical support is already stronger than many accepted CV works of similar scope. The two major points are genuine but addressable within a revision cycle; I do not see a load-bearing internal contradiction. Fit for a top ML/CV venue is reasonable provided the operating envelope of the weak-refraction assumption is clarified."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The one thing worth knowing is that GHOST is a practical, well-engineered preprocessor that turns transparent regions into opaque, structure-preserving RGB so that off-the-shelf models (DA3, MoGe2, VGGT, DUSt3R, π3, etc.) suddenly work without any retraining. Tables 2–4 and the zero-shot TROS numbers show large, consistent gains; the ablations (Tables 5–6) correctly show that partial pipelines make things worse. That is the real contribution.\n\nWhat is new is not any single module—DINOv3 features, SPADE, gated convs, alpha matting, and normal estimation all exist—but the specific four-stage cascade (TransDINO → TransDecomp → DAF-Net → GeoSemTransNet) that injects geometric curvature into the texture hallucination. The weak-refraction alpha-blending assumption (Snell’s-law argument, Eq. 7) is stated explicitly and is reasonable for thin-walled objects; the paper notes that strong refraction simply lowers alpha and falls back to surface synthesis. No hidden circularity: losses are standard, downstream models stay frozen, and evaluation is on public data.\n\nSoft spots are real but proportionate. No code, no error bars, no statistical tests, and free parameters (loss weights, LR schedule) are not swept. The weak-refraction premise is untested on thick or high-index glass, and the invented module names are a bit heavy. None of these sink the central claim.\n\nThis is for people who actually need transparent-object depth or reconstruction in robotics or 3D vision and do not want to retrain foundation models. It is a systems paper, not a theory paper. I would send it to peer review; the experimental design is thorough enough that a serious referee can decide. Worth reading if you work on the problem; cite it if you need a drop-in preprocessor.","headline":"Solid systems paper: a four-stage geometry-guided preprocessor that actually unlocks frozen SOTA depth/reconstruction models on transparent objects, with thorough multi-model evidence and honest ablations.","tokens_in":17066,"tokens_out":484,"would_cite":true,"duration_ms":5698,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"A preprocessing pipeline turns transparent objects into opaque, geometry-preserving RGB so ordinary depth and 3D models work without retraining.","keywords":["transparent objects","depth estimation","3D reconstruction","surface normals","alpha matting","geometry-guided synthesis","visual foundation models","preprocessing"],"falsifier":"On a dataset of thick, high-refraction transparent objects where the weak-refraction assumption fails, measure whether depth and reconstruction metrics after GHOST still improve over the raw-image baseline; if they do not, the central claim collapses.","tokens_in":17056,"feed_emoji":"🔍","tokens_out":589,"duration_ms":9409,"temperature":0.7,"pith_summary":"Transparent surfaces break the diffuse-reflection assumptions that almost every modern depth estimator and feed-forward 3D reconstructor rely on, so those models routinely return background geometry instead of the glass surface itself. GHOST is a four-stage preprocessing stack that first isolates the transparent region, then recovers an alpha matte, a foreground intensity map and a surface-normal field, and finally synthesizes a fully opaque RGB image whose textures still obey the recovered 3D shape. Because the output is still a plain RGB photograph, any off-the-shelf monocular depth or multi-view reconstruction network can be applied unchanged. Experiments on ClearGrasp, ClearPose and related benchmarks show large gains in depth accuracy, normal angular error and reconstruction completeness for several leading foundation models. The practical claim is that one geometry-guided “opaquification” step can restore the utility of the entire existing RGB-centric 3D vision stack for transparent objects.","feed_headline":"Transparent glass becomes opaque RGB so depth models work","feed_subtitle":"One geometry-guided preprocess restores accuracy for off-the-shelf 3D estimators without any retraining","key_machinery":"GHOST, a four-module cascade (TransDINO for masks, TransDecomp for alpha and foreground, DAF-Net for normals, GeoSemTransNet for SPADE-conditioned texture synthesis) that turns transparent pixels into structurally consistent opaque RGB.","core_discovery":"Geometry-guided hallucination of opaque surface textures—recovering mask, alpha, foreground and normals, then synthesizing an RGB image that looks opaque yet preserves the original 3D structure—is sufficient to let unmodified state-of-the-art depth and reconstruction models produce accurate geometry on transparent objects.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["Geometry cues hallucinate opaque textures that fix transparent depth","GHOST turns glass into structure-preserving opaque RGB for 3D models","Recover masks normals then synthesize opaque RGB transparent objects keep shape","Geometry-guided preprocess lets stock depth nets handle transparent glass","Opaque surface textures restore accurate geometry on transparent objects"],"cache_read_input_tokens":128,"weakest_assumption_plain":"The method assumes that light transport through thin transparent objects can be treated as simple alpha blending of a foreground layer with an inpainted background, which holds only when refraction is weak.","fun_headline_variants_meta":{"raw":{"variants":["Geometry cues hallucinate opaque textures that fix transparent depth","GHOST turns glass into structure-preserving opaque RGB for 3D models","Recover masks normals then synthesize opaque RGB transparent objects keep shape","Geometry-guided preprocess lets stock depth nets handle transparent glass","Opaque surface textures restore accurate geometry on transparent objects"]},"model":"grok-4.5","effort":"low","cost_usd":0.002974,"raw_usage":{"total_tokens":1031,"prompt_tokens":707,"num_sources_used":0,"completion_tokens":84,"cost_in_usd_ticks":29740000,"prompt_tokens_details":{"text_tokens":707,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":240,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":707,"tokens_out":84,"duration_ms":2857,"temperature":1.0,"reasoning_tokens":240,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-14T06:55:53.352714+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"On a dataset of thick, high-refraction transparent objects where the weak-refraction assumption fails, measure whether depth and reconstruction metrics after GHOST still improve over the raw-image baseline; if they do not, the central claim collapses.","supporting_citations":[],"review_version":1}