{"id":"00587c36-9bf0-4a2e-8cff-d6a6c61f78f1","arxiv_id":"2504.14221","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new 8,450-sample multimodal industrial anomaly detection dataset adds pseudo-3D photometric stereo data to RGB and high-resolution point clouds, with a fusion benchmark that improves average detection accuracy.","lead":"Real-IAD D3 introduces a multimodal industrial anomaly detection dataset with RGB images, micrometer-level 3D point clouds, and a new pseudo-3D modality from photometric stereo across 20 product categories. The paper also proposes a fusion baseline (D3M) that combines all three modalities and reports improved detection accuracy on average.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Pseudo-3D modality unvalidated: photometric stereo assumes Lambertian surfaces and known light directions on glossy materials, and reported ablations do not isolate pseudo-3D's contribution; per-category results also contradict 'consistently surpasses'.","rationale":"The reader's weakest assumption (Lambertian photometric stereo) is sound and matches my primary concern, but I would frame the load-bearing issue as evidential: the pseudo-3D modality is never validated against the dataset's own 3D geometry, and the experimental design does not control for architecture when claiming the modality's benefit. The paper's average results are real, and a large, precisely aligned multimodal dataset would be valuable even if the pseudo-3D maps are not physically accurate normals; hence I do not move to REJECT. The dataset URL and code release are secondary but still block verification. The strongest check is a per-material normal-accuracy comparison using the released point clouds; this would either vindicate the pseudo-3D construction or force a re-scoped claim. Since the reader already returned CONDITIONAL and my concern tightens the condition without changing the disposition, the verdict stays UNCHANGED.","tokens_in":14584,"tokens_out":7045,"duration_ms":60299,"concrete_test":"Quantitatively validate the pseudo-3D normals: from the released high-resolution point clouds, estimate surface normals by local plane fitting at each pixel and compare with the released photometric-stereo normals on all 20 categories; report per-category mean angular error and the error restricted to defect regions. If specular categories (ferrite_bead, power_jack, headphone_jack_socket) show mean errors above about 20 degrees while matte categories are below about 10 degrees, the Lambertian assumption is violated exactly where the pseudo-3D benefit is claimed, and the modality should be re-labeled or recalibrated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central novelty of Real-IAD D³ is the pseudo-3D modality, but the paper provides no evidence that the photometric-stereo reconstruction in Sec. 3.1 is actually valid on this material set. The model I = L·n and n = (L^T L)^{-1} L^T I assumes Lambertian reflectance, known fixed light directions, and no shadows or interreflections. Many of the 20 categories are glossy metals and plastics (ferrite_bead, power_jack, headphone_jack_socket, dc_power_connector), where specularities violate the model. The paper does not report light-direction calibration, albedo normalization, or a comparison of recovered normals against the micrometer-level 3D point clouds it also provides, so the 'pseudo-3D depth' could be a materially corrupted signal rather than surface orientation. The downstream evidence is also not controlled: Table 3 shows PatchCore(2D+Pseudo-3D) improves I-AUROC over PatchCore(2D) (0.839 vs 0.747) but lowers P-AUROC (0.867 vs 0.875), and the final D³M gain relative to M3DM comes with a different fusion architecture and training, so it does not isolate the pseudo-3D contribution. Per-category numbers in Table 2 further contradict 'consistently surpasses': D³M is worse than M3DM on P-AUROC in 9 of 20 categories, e.g., connector_housing_female 0.951 vs 0.979, and miniature_lifting_motor I-AUROC 0.823 vs 0.975. Without validation of the pseudo-3D modality itself, the dataset's headline novelty and the method's attribution of gains to that modality remain unestablished.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces Real-IAD D3, a multimodal industrial anomaly detection dataset containing 8,450 samples across 20 product categories, with synchronized high-resolution RGB images, micrometer-level 3D point clouds (precision 0.002 mm), and a pseudo-3D modality computed via four-light photometric stereo. The authors also propose D3M, a benchmark method that fuses RGB, pseudo-3D, and point-cloud features using channel-spatial swapping, unsupervised contrastive alignment, and multi-bank decision fusion. Experiments report average I-AUROC 0.890 and P-AUROC 0.937 for D3M on the proposed dataset, outperforming single-modality and 2D+3D baselines on average. The paper also includes resolution-downsampling and feature-interpolation analyses to highlight the value of high-resolution point clouds.","tokens_in":14967,"tokens_out":5295,"duration_ms":41303,"significance":"If the dataset is released as described, it would be a substantial resource for the IAD community: it is larger and higher-resolution than MVTec 3D-AD and Real3D-AD, covers 69 defect groups across diverse real-world materials, and provides synchronized 2D/pseudo-3D/3D data with public access. The pseudo-3D modality is a novel contribution that could aid detection of subtle surface defects, though its validity on glossy and textured materials is not yet demonstrated. The D3M benchmark is a useful baseline but is largely composed of existing components (DINO, PointMAE, M3DM-style contrastive fusion and decision layers); its methodological novelty is limited. The average performance gains over M3DM are real, but the paper's stronger claim of 'consistently surpassing' all 2D+3D methods is contradicted by the per-category numbers, and the ablations do not isolate the pseudo-3D contribution. The paper would become publishable after a major revision that validates the pseudo-3D modality and corrects or qualifies the performance claims.","major_comments":[{"comment":"The claim that D3M 'consistently surpasses' 2D+3D methods is contradicted by the per-category results. For example, on miniature_lifting_motor the I-AUROC drops from 0.975 (M3DM) to 0.823 (D3M), and on power_jack it drops from 0.996 to 0.973. In addition, D3M is worse than M3DM on P-AUROC in at least eight categories (connector_housing_female 0.951 vs 0.979, dc_power_connector 0.947 vs 0.950, ethernet_connector 0.970 vs 0.978, fork_crimp_terminal 0.946 vs 0.964, fuse_holder 0.915 vs 0.948, lattice_block_plug 0.941 vs 0.958, miniature_lifting_motor 0.961 vs 0.991, purple_clay_pot 0.922 vs 0.953). The paper should be revised to state that D3M achieves the best average performance, not consistent per-category superiority, and the authors should discuss the categories where the additional pseudo-3D modality degrades performance.","section":"§5.1, Table 2"},{"comment":"The experimental design does not isolate the contribution of the pseudo-3D modality. D3M differs from M3DM in several architectural components (CSS swapping, contrastive fusion with swapped 2D features, and a triple memory bank), so the improvement of D3M over M3DM in Table 2 cannot be attributed to the pseudo-3D input alone. Moreover, Table 3 reports M3DM(2D+3D) with I-AUROC 0.752 and P-AUROC 0.687, which is inconsistent with the M3DM averages in Table 2 (0.841 I-AUROC, 0.922 P-AUROC) and is not explained in the text. The authors should provide a controlled ablation that runs the same architecture with and without the pseudo-3D modality, and they must reconcile the M3DM numbers between Table 2 and Table 3.","section":"§5.2, Table 3"},{"comment":"The pseudo-3D modality is generated by solving n = (L^T L)^{-1} L^T I under the assumption of Lambertian reflectance, known fixed light directions, and no shadows or interreflections. Many of the 20 categories are glossy metals and plastics (ferrite_bead, power_jack, headphone_jack_socket, dc_power_connector), where specularities violate the Lambertian assumption. The paper provides no calibration details for the light directions, no albedo normalization, and, critically, no validation of the recovered normal maps against the co-captured micrometer-level point clouds. Without such validation, the pseudo-3D modality could be a corrupted signal, and the headline claim that this modality improves detection is not established. I request that the authors add either a quantitative comparison of the photometric-stereo normals against the 3D scans on representative categories or an ablation in which the pseudo-3D channel is replaced by a noise or constant signal, to demonstrate that the performance gain is specifically due to meaningful surface-orientation information.","section":"§3.1, photometric stereo equation"},{"comment":"The paper interchangeably calls the photometric-stereo output 'pseudo-3D depth' and 'pseudo-3D surface normals', but Eq. (1) and Eq. (2) estimate a surface normal field, not depth. This is a conceptual mismatch that affects how readers and downstream methods interpret the modality. The authors should consistently describe the modality as surface normals (or, if they truly intend depth, explain how depth is integrated from normals) and adjust the text, captions, and figures accordingly.","section":"Throughout (Sec. 3.1, Fig. 2, Fig. 3)"}],"minor_comments":[{"comment":"There are several typos and inconsistencies: 'Pesudo-3D' in Table 3 and Figure 7, 'Tabel' in Table 3 caption, and inconsistent category names (e.g., 'common-mode-filter' vs 'commonmodefilter', 'connector_housing-female' vs 'connectorhousing-female'). These should be standardized.","section":"Throughout"},{"comment":"The URL in the abstract contains a space ('realiad4ad.github.io/Real-IAD D3'); it should be corrected to a proper link, e.g., 'https://realiad4ad.github.io/Real-IAD-D3'.","section":"Abstract and Sec. 1"},{"comment":"The table lists '3D Point-Cloud Resolution' as 0.01 mm and 'Point Precision' as 0.002 mm, but the text in Sec. 3 says 'accuracy of 0.002 mm' and 'four-point cloud resolutions'. The relationship between resolution, precision, and accuracy should be defined clearly to avoid confusion.","section":"Table 1"},{"comment":"Table 4 reports PointMAE at Original vs 4x resolution but does not include the 40x condition shown in Fig. 8; the text and figure should be consistent about which downsampling factors are evaluated quantitatively.","section":"Sec. 5.3, Table 4"},{"comment":"The figure labels 'Zoom in=15', 'Zoom in=10', and 'Zoom in=3' are not explained in the text or caption; please clarify what these values represent.","section":"Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The paper's dataset contribution is potentially valuable, but the central claims about the pseudo-3D modality and the 'consistently surpasses' wording require substantial revision. The inconsistency between Tables 2 and 3 for the M3DM baseline is a red flag that the evaluation protocol may not be controlled; the authors must address this before the paper can be accepted. If the authors can add a validation of the pseudo-3D normals and a proper ablation isolating the modality, the paper would likely be suitable for publication in a datasets-and-benchmarks venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Real-IAD D3 is worth knowing about even though it needs work. The dataset is the real contribution: RGB, pseudo-3D (photometric-stereo normals), and high-res point clouds (0.002 mm) for 20 categories, 8,450 samples, synchronized on one rig. That scale and resolution beat MVTec 3D-AD and Real3D-AD, and the pseudo-3D modality is new in this area. The downsampling study (Table 4) is a nice demonstration of why high-res point clouds matter for small parts. If the dataset is actually released and the link fixed, it will be a useful benchmark.\n\nThe soft spots are mostly in the evaluation and claims. The photometric-stereo step is pure Lambertian least squares (Sec. 3.1), and many of your categories are glossy metal and plastic parts. The paper doesn't calibrate the light directions, doesn't handle specularities, and doesn't compare the recovered normals against the point clouds it also provides. So we don't know whether the pseudo-3D is surface orientation or just a shading artifact. The downstream numbers don't settle it: Table 3 shows PatchCore(2D+pseudo-3D) improves I-AUROC over PatchCore(2D) but worsens P-AUROC (0.867 vs 0.875). And D3M's average gain over M3DM comes with a different fusion module, so it doesn't isolate pseudo-3D.\n\nThe claim that D3M 'consistently surpasses' 2D+3D methods is simply wrong on the paper's own Table 2: M3DM beats D3M on miniature_lifting_motor I-AUROC (0.975 vs 0.823), on power_jack (0.996 vs 0.973), and on 9 of 20 categories in P-AUROC. That kind of overstatement in the text should be fixed. Also missing code, a URL with a space, and no error bars.\n\nNone of this kills the dataset. The central resource is plausible and likely useful, and the concerns are addressable. I'd send this to peer review and ask for validation of the pseudo-3D modality, per-category numbers, code with a commit hash, and softened claims. The method contribution is incremental (M3DM plus 10% feature swapping), but as a baseline it's fine.","headline":"Real-IAD D3 is a potentially valuable dataset with a genuinely new pseudo-3D modality, but the paper overclaims the method's consistency and never validates the photometric-stereo output.","tokens_in":15582,"tokens_out":3074,"would_cite":true,"duration_ms":25971,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Real-IAD D3 adds photometric-stereo pseudo-3D views to RGB and micrometer point clouds and reports that the three-modality fusion beats every 2D-only and 2D+3D baseline on average.","keywords":["industrial anomaly detection","multimodal dataset","photometric stereo","pseudo-3D","point cloud","anomaly localization","D3M benchmark","Real-IAD D3"],"falsifier":"Take the glossy categories (for example ferrite bead, dc power connector, or knob cap), compute the photometric-stereo normal maps, and compare them against normals derived from the 0.002 mm point clouds; if the pseudo-3D normals deviate systematically in specular regions, rerun the D3M benchmark on only the matte categories and on only the glossy categories. The central claim would be settled by showing whether the pseudo-3D gain is present in the glossy subset or comes entirely from diffuse materials.","tokens_in":14379,"feed_emoji":"🔍","tokens_out":8488,"duration_ms":70362,"temperature":0.7,"pith_summary":"The paper introduces Real-IAD D3, a multimodal industrial anomaly detection dataset with 8,450 samples across 20 product categories, each sample carrying synchronized high-resolution RGB images, micrometer-level point clouds (down to 0.002 mm precision), and a pseudo-3D depth/normal view computed from four photometric-stereo light directions. The central claim is that this third modality supplies surface-orientation information that 2D color alone and 2D-plus-3D combinations miss, especially for fine scratches, pits, and dents on small components. To demonstrate the claim, the paper builds D3M, a fusion benchmark that swaps 10% of RGB and pseudo-3D features, aligns 2D and 3D features contrastively, and scores anomalies against memory banks. On average D3M reaches 0.890 image-level AUROC and 0.937 pixel-level AUROC, above 2D-only, 3D-only, and 2D+3D baselines, while experiments with downsampled point clouds show fine defects vanish at coarser resolution. The dataset and code are released publicly so other groups can test multimodal anomaly detection under realistic industrial conditions.","feed_headline":"Pseudo-3D views lift industrial defect detection to 0.890","feed_subtitle":"Real-IAD D3 pairs RGB and micrometer point clouds with photometric stereo, beating 2D+3D baselines.","key_machinery":"The load-bearing mechanism is photometric stereo: four directional light sources illuminate each part, and the pixel intensities $I(x,y)$ are converted to surface normals by solving $n(x,y) = (L^\\top L)^{-1} L^\\top I(x,y)$, where $L$ is the known light-direction matrix. The result is a pseudo-3D image of surface orientation that encodes shallow defects as sudden normal changes. On top of this, the D3M benchmark adds a Channel-Spatial Swapping module that exchanges 10% of RGB and pseudo-3D feature-map channels and spatial blocks, an unsupervised contrastive loss that aligns swapped 2D patches with point-cloud patches, and three memory banks scored by one-class SVMs to produce image-level and pixel-level anomaly maps. The 3D point clouds are captured on the same platform with a four-direction DLP structured-light system, which the paper credits for the strict RGB/3D/pseudo-3D alignment.","core_discovery":"Real-IAD D3's core discovery is that a pseudo-3D modality obtained by photometric stereo is a usable third signal for unsupervised industrial anomaly detection, not just a substitute for measured 3D. The dataset contains 5,000 normal and 3,450 anomalous samples across 20 categories with 69 defect groups, covers defect areas as small as 0.46% of the part, and captures point clouds of up to 16.2 million points at 0.002 mm accuracy. The authors argue that pseudo-3D surface normals reveal shallow texture and orientation defects that are hard to see in RGB and too small for coarser point clouds, and their D3M benchmark backs this with average I-AUROC 0.890 and P-AUROC 0.937, outperforming the 2D+3D M3DM baseline and a 2D+pseudo-3D PatchCore combination. In their framing, strictly aligned multimodal capture on a single platform is what makes the fusion possible, and the additional modality raises the ceiling for fine-defect localization.","pith_inferences":["One consequence the paper leaves implicit is that the pseudo-3D channel inherits photometric stereo's diffuse-reflection assumption; on shiny metal and plastic parts the normal equation can produce artifacts, so a material-wise breakdown of the 0.890 average could reveal categories where the third modality hurts.","Because RGB, pseudo-3D, and point clouds are captured by one fixed rig, part of the fusion gain may be specific to that alignment; running D3M on a different acquisition setup would test whether the benefit transfers.","The near-parity of the 2D+pseudo-3D result with the full three-modality result suggests measured 3D pays off mainly on larger geometric defects, so a per-defect-type analysis could tell practitioners when micrometer point-cloud hardware is worth the cost."],"forward_implications":["Photometric-stereo pseudo-3D should be treated as a complementary third channel rather than a replacement for scanned 3D, since the full three-modality fusion gives the best scores.","The 2D+pseudo-3D combination already shows strong average scores in the paper's PatchCore-style comparison, so adding a four-light capture stage may yield a large part of the fusion gain even before measured 3D is included.","Resolution is decisive for small components: downsampling point clouds by 4x or 40x degrades detection of scratches and pits, so micrometer-level capture is necessary rather than optional in this regime.","The high native point-cloud density removes the need for post-backbone feature interpolation on Real-IAD D3, whereas on MVTec 3D-AD pixel-AUROC drops by 0.111 without it; practitioners can skip interpolation when raw resolution is high."],"supporting_citations":[{"why":"The RGB+3D benchmark that Real-IAD D3 extends and compares against in scale, resolution, and modality coverage.","marker":"[4]"},{"why":"The large-scale point-cloud anomaly detection dataset used to show Real-IAD D3's higher point precision and richer defect set.","marker":"[19]"},{"why":"The real-world multi-view 2D dataset that precedes Real-IAD D3 and supplies the industrial part categories and defect-making experience.","marker":"[28]"},{"why":"The 2D+3D hybrid-fusion method whose contrastive fusion and memory banks D3M adapts, and whose average performance is the direct baseline to beat.","marker":"[30]"},{"why":"The memory-bank anomaly detector used as the representative 2D and 2D+pseudo-3D baseline in the modality-combination experiments.","marker":"[21]"},{"why":"The standard 2D anomaly detection dataset that motivates why geometric depth modalities are needed.","marker":"[2]"},{"why":"The synthetic multimodal dataset with RGB, depth, and normal maps that anchors the related-work comparison for multimodal IAD data.","marker":"[5]"},{"why":"The depth-simulation method that motivates using synthetic or pseudo depth cues to recover subtle 3D surface defects.","marker":"[33]"}],"fun_headline_variants":["Three modalities outdo two: Real-IAD D3 hits 0.890 AUROC","Pseudo-3D modality boosts anomaly detection to 0.890 I-AUROC","Real-IAD D3: photometric stereo supplies the third signal for defect detection","Photometric stereo gives anomaly detection a third signal lifting AUROC to 0.890"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The pseudo-3D views are computed by assuming each surface scatters light evenly from a known direction, which is false for shiny metal and plastic parts; if that assumption breaks on enough categories, the third modality adds noise rather than signal and the reported gain over 2D+3D would not survive.","fun_headline_variants_meta":{"raw":{"variants":["Three modalities outdo two: Real-IAD D3 hits 0.890 AUROC","Pseudo-3D modality boosts anomaly detection to 0.890 I-AUROC","Real-IAD D3: photometric stereo supplies the third signal for defect detection","Photometric stereo gives anomaly detection a third signal lifting AUROC to 0.890"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001704,"raw_usage":{"total_tokens":6771,"prompt_tokens":994,"completion_tokens":5777,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":610,"completion_tokens_details":{"reasoning_tokens":5685}},"tokens_in":610,"tokens_out":5777,"duration_ms":35170,"temperature":1.0,"reasoning_tokens":5685,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:52:51.848628+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take the glossy categories (for example ferrite bead, dc power connector, or knob cap), compute the photometric-stereo normal maps, and compare them against normals derived from the 0.002 mm point clouds; if the pseudo-3D normals deviate systematically in specular regions, rerun the D3M benchmark on only the matte categories and on only the glossy categories. The central claim would be settled by showing whether the pseudo-3D gain is present in the glossy subset or comes entirely from diffuse materials.","supporting_citations":[{"cited_title":"Real3d- ad: A dataset of point cloud anomaly detection","cited_arxiv_id":null,"evidence_quote":"The large-scale point-cloud anomaly detection dataset used to show Real-IAD D3's higher point precision and richer defect set."},{"cited_title":"Multimodal industrial anomaly detection via hybrid fusion","cited_arxiv_id":null,"evidence_quote":"The 2D+3D hybrid-fusion method whose contrastive fusion and memory banks D3M adapts, and whose average performance is the direct baseline to beat."},{"cited_title":"Towards to- tal recall in industrial anomaly detection","cited_arxiv_id":null,"evidence_quote":"The memory-bank anomaly detector used as the representative 2D and 2D+pseudo-3D baseline in the modality-combination experiments."},{"cited_title":"Mvtec ad–a comprehensive real-world dataset for unsupervised anomaly detection","cited_arxiv_id":null,"evidence_quote":"The standard 2D anomaly detection dataset that motivates why geometric depth modalities are needed."},{"cited_title":"The eyecandies dataset for unsupervised multimodal anomaly detection and local- ization","cited_arxiv_id":null,"evidence_quote":"The synthetic multimodal dataset with RGB, depth, and normal maps that anchors the related-work comparison for multimodal IAD data."},{"cited_title":"Cheat- ing depth: Enhancing 3d surface anomaly detection via depth simulation","cited_arxiv_id":null,"evidence_quote":"The depth-simulation method that motivates using synthetic or pseudo depth cues to recover subtle 3D surface defects."}],"review_version":1}