{"id":"d4a59804-3179-415d-875d-7a9fbfbdcbed","arxiv_id":"1908.02893","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Canny edges from an RGB image are projected into 3D, encoded with flipped TSDF, and fused with depth in a U-Net based CNN, improving semantic scene completion by about 3 points on SUNCG while matching complex two-stage methods on NYU.","lead":"EdgeNet takes a single color-and-depth photo of a room and predicts the complete 3D layout, including hidden walls and furniture with semantic labels. It is a simpler one-stage method that adds object outlines from the photo, improving detection of flat objects such as posters and TVs.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Edge benefit rests on single-run gaps of +0.5/+0.8 IoU that could be training noise; multi-seed ablation is needed.","rationale":"The reader's weakest assumption concerned cross-paper evaluation compatibility, which is a real threat to the SOTA claims. However, the more fundamental threat to the paper's central contribution is internal: the ablations that isolate the effect of edges are small, unreplicated, and confounded with architecture and training-pipeline changes. The reader's CONDITIONAL verdict remains appropriate, but the condition should explicitly require multi-seed evidence for the edge benefit, not only release of code and alignment of SOTA claims. I therefore keep the verdict unchanged while identifying a different, arguably more load-bearing weakness.","tokens_in":12250,"tokens_out":3353,"duration_ms":36853,"concrete_test":"Run EdgeNet-D, EdgeNet-EF, and EdgeNet-MF, plus SSCNet* and SSCNet-E as controls, with at least 5 random seeds on SUNCG and on the SUNCG-to-NYUDv2 fine-tuning protocol. Report mean and standard deviation of average IoU, and the paired per-scene difference EF−D and MF−D with a 95% confidence interval. Also include an EdgeNet-EF control where the edge volume is replaced by random binary noise matched for sparsity. If the confidence interval for EF−D includes 0, or the noise control yields the same gain as real edges, the claimed benefit of RGB edge encoding is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that F-TSDF-encoded Canny edges provide a genuine RGB signal for 3D semantic scene completion, making EdgeNet the first end-to-end method to beat depth-only. The evidence for this is the gap between paired ablations: SSCNet-E vs SSCNet* (65.7 vs 65.2, +0.5) and EdgeNet-EF vs EdgeNet-D (70.3 vs 69.5, +0.8) on SUNCG. These gaps are reported for a single run, with no error bars, and the NYU results are inconsistent: no fusion scheme consistently beats depth-only or TNetFuse. The much larger gains in Table I come from the new architecture and training pipeline (SSCNet* 65.2 vs original SSCNet 46.4; EdgeNet-D 69.5), so the 'edge' contribution is confounded with model capacity, regularization, and optimization changes. If the +0.5/+0.8 gaps vanish under repeated runs, the paper's main novelty—edge F-TSDF fusion—is not established, even if the architecture and training pipeline are strong. This is more load-bearing than the cross-paper comparison issue because it attacks the method's core contribution, not just the SOTA ranking.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"EdgeNet proposes a semantic scene completion method from a single RGB-D image. It extracts Canny edges from the RGB image, projects them to 3D using the depth map, voxelizes them, and encodes both the depth surface and the edge points with flipped TSDF (F-TSDF). The two volumes are then fed into a U-Net-like 3D CNN built from ResNet modules with dilated convolutions, and the paper evaluates three fusion schemes: early (EdgeNet-EF), middle (EdgeNet-MF), and late (EdgeNet-LF). Training uses offline F-TSDF preprocessing, weighted categorical cross-entropy, and a one-cycle learning-rate schedule. On SUNCG, EdgeNet-EF reports 70.3 average semantic IoU, above the reported VVNetR-120 value of 66.7; on NYUDv2, EdgeNet-MF reports 33.7 after fine-tuning on SUNCG plus NYU, close to the reported TNetFuse value of 34.4. The paper also reports ablations isolating the training pipeline (SSCNet*), the architecture (EdgeNet-D), and the edge input (SSCNet-E and the EdgeNet fusion variants).","tokens_in":12492,"tokens_out":6842,"duration_ms":71934,"significance":"If the reported gains are reproducible, the paper makes a useful contribution: binary edge volumes encoded with F-TSDF provide a simple and transferable RGB cue for 3D semantic scene completion without requiring a separately trained 2D segmentation network. The ablation design is a genuine strength: SSCNet* and EdgeNet-D isolate the training-pipeline and architecture contributions from the edge contribution, and the three fusion schemes explore a well-defined design space while keeping memory constant. However, the central claim that edges consistently improve depth-only performance is currently supported by single-run differences smaller than one IoU point on SUNCG, there is no depth-only EdgeNet baseline on NYU, and the state-of-the-art comparisons are made with numbers extracted from other papers under unspecified evaluation protocols. If the robustness and comparability issues are addressed, the contribution would be significant for the semantic scene completion community.","major_comments":[{"comment":"The evidence that F-TSDF encoded edges improve semantic scene completion rests on single training runs. On SUNCG the paired edge-vs-depth gaps are +0.5 (SSCNet-E vs SSCNet*) and +0.8 (EdgeNet-EF vs EdgeNet-D), while EdgeNet-LF is actually 2.7 points below EdgeNet-D (66.8 vs 69.5). These margins are comparable to typical stochasticity in 3D CNN training, and no variance or number of seeds is reported. The authors should provide mean and standard deviation over at least three runs for at least the paired ablations, or otherwise demonstrate that the edge advantage is not due to chance.","section":"IV-D1, Table I"},{"comment":"On NYUDv2 the paper reports only EdgeNet variants with depth and edges; there is no EdgeNet-D depth-only baseline trained under the same protocols. The claim that edges help on real data therefore rests on comparisons to the original SSCNet, which differs in architecture and training pipeline. Without an NYU-trained EdgeNet-D, the edge contribution on real data is confounded with architecture and pipeline changes. The authors should add the depth-only EdgeNet baseline on NYU.","section":"IV-E, Table II"},{"comment":"The SUNCG comparisons to previous methods are not apples-to-apples. The paper states that because the original SUNCG train/test sets did not include RGB images, the authors re-rendered new depth and RGB images and regenerated ground-truth volumes from the scene meshes. The numbers for SSCNet, DCRF, and VVNetR-120 in Table I were, however, extracted from the original papers, which used a different rendering and possibly different ground-truth generation. The large improvement of SSCNet* over the reported SSCNet (65.2 vs 46.4) may therefore be partly an artifact of the new data, not only of the new training pipeline. The authors should either evaluate prior methods on the same re-rendered data or provide evidence that their renderings and regenerated volumes reproduce the original evaluation protocol.","section":"IV-A and IV-D, Table I"},{"comment":"The real-data comparisons are made with numbers 'extracted from the original papers' without shared evaluation code or a verified common ground-truth voxelization. The paper uses ground truth based on [20] and [21], but it does not establish that prior work used the same voxelization, test split, and evaluation script. Since the paper's stated contribution includes achieving state-of-the-art or equivalent performance on NYUDv2, this is load-bearing: small numerical differences such as 33.7 vs 34.4 cannot be interpreted as equivalence or near-equivalence unless the evaluation protocol is identical. The authors should release their evaluation code and ground-truth generation pipeline, or run official implementations of prior methods under that same pipeline.","section":"IV-D, Table II"}],"minor_comments":[{"comment":"The text refers to 'SSNet*' in the description of the baseline retrained with the new pipeline, but Table I and the rest of the paper use 'SSCNet*'; the notation should be made consistent.","section":"IV-D1"},{"comment":"Figure 4 shows the input volume as '240x144x200' in the network diagram, while Section III-C and the rest of the text specify a 240x144x240 volume; please correct the inconsistency.","section":"III-B, Figure 4"},{"comment":"The abstract claims an improvement of '6.9% over the state-of-the-art result on real data, for end-to-end approaches,' but no table directly supports this specific figure; please specify the exact comparison and how it is computed, or rephrase the claim to match the reported numbers.","section":"Abstract and Section IV-E"},{"comment":"The description of the loss-weighting scheme is difficult to follow: the definition of the ratio r and the construction of the random tensor randoccl should be clarified, including how the weights are normalized across the batch, since this affects all reported results.","section":"III-B"},{"comment":"There is a repeated typo 'approachs' in the first paragraph of the Discussion, and Section IV-D1 uses 'Ednet' instead of 'EdgeNet'; these should be corrected.","section":"V"}],"recommendation":"major_revision","confidential_remarks":"The core idea—F-TSDF encoded projected edges as an RGB signal for semantic scene completion—is interesting and potentially publishable, and the ablation structure is well organized. The main concern is robustness: the paper's central claim about the benefit of edges is based on single-run differences of less than one IoU point on SUNCG, and there is no depth-only EdgeNet baseline on NYU. In addition, the SUNCG re-rendering issue may be more serious than the authors appear to recognize: comparisons to numbers from the original SUNCG experiments are only valid if the re-rendered depth and regenerated ground truth match the original evaluation protocol exactly. I would encourage the editor to request at minimum multi-seed ablations and a clear accounting of the SUNCG data generation and evaluation comparability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is worth engaging, but the headline edge benefit is not yet proven. What's new: F-TSDF encoding of Canny edges projected into 3D, fused with depth in a U-Net with ResNet blocks and three fusion schemes. That is a genuinely new representation and a solid engineering contribution, and the paper does a good job isolating the effects of its training pipeline, architecture, and edge signal through clean ablations.\n\nWhat it does well: the ablations are well structured. SSCNet* vs SSCNet shows the training pipeline accounts for most of the improvement over prior depth-only results. EdgeNet-D shows the U-Net/ResNet architecture helps on its own. The idea of applying F-TSDF to binary edge volumes is sensible, and the qualitative results, especially for wall-mounted objects like pictures, make an intuitive case for why edges could help. The paper is also honest about NYU ground truth problems.\n\nSoft spots, in proportion: the central new claim, that edges improve depth-only SSC, rests on SUNCG gains of +0.5 IoU (SSCNet-E vs SSCNet*) and +0.8 IoU (EdgeNet-EF vs EdgeNet-D), each from a single run with no error bars. On NYU, no fusion scheme consistently beats depth-only. The stress-test note is right: this is load-bearing, because the larger gains come from the architecture and pipeline, not from the edge signal. The comparisons to prior work are also cross-paper numbers, extracted from other papers with no shared evaluation code or ground truth version verified. The abstract's claim of 6.9% improvement over state-of-the-art on real data does not obviously match the tables (EdgeNet-MF is 33.7 vs TNetFuse's 34.4 on SUNCG+NYU). No code or data is released, though the paper says the pipeline will be made public.\n\nThese are addressable weaknesses. The architecture and training pipeline are solid contributions on their own, and the edge representation is plausible. But the paper's signature claim, that this is the first end-to-end approach to use RGB to improve over depth-only, is not yet established.\n\nWho this is for: researchers working on 3D semantic scene completion, indoor robotics, and AR. The paper deserves a serious referee, but the verdict should be conditional on multi-seed ablations and artifact release. I would accept it for review and ask for repeated runs and a clearer mapping between abstract claims and tables.","headline":"EdgeNet makes a plausible architectural contribution, but its central claim that F-TSDF-encoded edges help rests on single-run gaps that need multi-seed verification.","tokens_in":13051,"tokens_out":1984,"would_cite":true,"duration_ms":20994,"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":"EdgeNet fuses RGB edges with depth to improve 3D semantic scene completion, reporting 70.3 average IoU on SUNCG and matching a two-stage rival on NYU.","keywords":["semantic scene completion","RGB-D fusion","Canny edge detection","flipped TSDF","3D convolutional network","SUNCG","NYUDv2","end-to-end training"],"falsifier":"Re-run VVNetR-120 and TNetFuse with EdgeNet's re-rendered SUNCG test images and its NYU ground-truth voxelisation and evaluation code. If they reproduce their reported scores, EdgeNet's 70.3 and 33.7 IoU results hold; if their scores move, the claimed gains are protocol artefacts. A further check is to feed EdgeNet a random binary volume with the same density as the projected edges: if the IoU gain over EdgeNet-D persists, the improvement is not specifically due to RGB edges.","tokens_in":12010,"feed_emoji":"🏠","tokens_out":7643,"duration_ms":71642,"temperature":0.7,"pith_summary":"EdgeNet is an end-to-end network for semantic scene completion that claims colour information can help predict complete 3D scenes when it is encoded as binary edge volumes rather than as raw RGB projections. The paper projects Canny edges from a single RGB-D image into 3D, applies flipped truncated signed distance (F-TSDF) to both the depth surface and the edge volume, and fuses the two inside a U-Net style 3D CNN. On SUNCG the best variant reaches 70.3 average IoU, 3.3 points above the strongest depth-only prior method, and on NYUDv2 it reaches 33.7 average IoU, within one point of a more complex two-stage colour fusion method. The practical stake is a simpler training pipeline: one network, no separate 2D segmentation pretraining, and lower memory use, with gains concentrated on flat objects like pictures and TVs that depth alone cannot distinguish.","feed_headline":"RGB edges push 3D scene completion past depth-only models","feed_subtitle":"Binary edge volumes, F-TSDF encoded, lift SUNCG semantic IoU to 70.3 and match a complex rival on NYU.","key_machinery":"The load-bearing object is the flipped truncated signed distance function (F-TSDF) applied to a projected Canny edge volume. F-TSDF is defined as $\\mathrm{sign}(\\mathrm{TSDF})\\cdot(1-|\\mathrm{TSDF}|)$, which turns the surface into a sharp discontinuity and gives dense gradients in the volume; the paper's move is to treat RGB edges as a second binary surface so that the same encoding can be applied to them. The second component is EdgeNet itself, a U-Net style 3D CNN with ResNet modules and dilated convolutions, evaluated in early, middle and late fusion schemes with a fixed channel budget so that memory use stays constant. The machinery works by giving the 3D CNN two complementary dense distance fields: one for geometry from depth and one for appearance boundaries from colour.","core_discovery":"The central claim is that previous attempts to add colour to semantic scene completion failed because colour is sparse after projection into a voxel grid, and that binarising appearance into edges before projection removes this sparsity. Because an edge volume is binary, the same F-TSDF encoding used for depth surfaces can be applied to it, producing a dense signed-distance field around projected edges. EdgeNet takes the depth F-TSDF volume and the edge F-TSDF volume, fuses them at one of three points in a 3D encoder-decoder built from ResNet modules with dilated convolutions, and trains the whole network with a batch-balanced categorical cross-entropy loss. The authors report that this is the first end-to-end approach in which RGB information improves semantic scene completion over depth only, and that the improvement is largest for classes such as TVs and generic objects.","pith_inferences":["The paper only tests Canny edges, but its transfer argument implies that any binary 3D cue that can be F-TSDF encoded, such as object boundaries, keypoints, or learned edge maps, could be fused the same way; that is an extension the authors did not run.","Because the paper shows that naive RGB projection fails while edge projection succeeds, a natural follow-up is to compare F-TSDF edge volumes against F-TSDF applied to dense semantic class probabilities, which would test whether the benefit comes from binarisation or from the edge cue specifically.","The NYU ground-truth errors the paper documents (missing windows, pictures, shelves) penalise accurate models, so a real-world benchmark with cleaner annotations would likely show a larger EdgeNet margin than the reported 33.7.","The fixed-memory fusion comparison suggests that late fusion loses accuracy because duplicated branches force channel cuts; a multi-scale attention mechanism could capture late-fusion benefits without that memory penalty."],"forward_implications":["On SUNCG, EdgeNet-EF reaches 70.3 average semantic IoU, surpassing the best depth-only method VVNetR-120 by 3.3 points.","Feeding F-TSDF edges into the older SSCNet architecture (SSCNet-E) improves over the depth-only SSCNet*, so the edge representation helps beyond EdgeNet itself.","On NYUDv2 with SUNCG pretraining and fine-tuning, EdgeNet-MF reaches 33.7 average IoU, within 0.7 points of TNetFuse while using a single end-to-end training stage.","The gains are concentrated in hard-to-detect classes such as TVs, pictures and generic objects, which are nearly invisible to depth after down-sampling.","The offline F-TSDF preprocessing cuts training time from 7 to 4 days on SUNCG and from 30 to 6 hours on NYUDv2 compared with the original SSCNet pipeline."],"supporting_citations":[{"why":"Defines the semantic scene completion task and the F-TSDF encoding that EdgeNet adopts for depth and edges.","marker":"[7]"},{"why":"Provides the Canny edge detector used to turn RGB images into binary edge maps before 3D projection.","marker":"[16]"},{"why":"Shows that naive projection of raw RGB channels into the voxel grid fails to improve completion, motivating the edge encoding.","marker":"[8]"},{"why":"Supplies TNetFuse and SNetFuse, the two-stage depth-plus-2D-segmentation approaches whose NYU results EdgeNet nearly matches.","marker":"[10]"},{"why":"Supplies VVNetR-120, the depth-only method that EdgeNet surpasses by 3.3 points on SUNCG.","marker":"[13]"},{"why":"Supplies the U-Net encoder-decoder design on which EdgeNet's architecture is based.","marker":"[17]"},{"why":"Supplies the ResNet modules used to build deeper 3D convolutional blocks with dilated convolutions.","marker":"[15]"},{"why":"Supplies the DCRF depth-only baseline reported in the comparison tables.","marker":"[12]"}],"fun_headline_variants":["Edge binarization turns sparse RGB into dense 3D cues","F-TSDF edge volumes densify RGB for 3D scene completion","EdgeNet fuses depth and edge F-TSDF volumes for semantics","Binarized edges solve colour sparsity in RGB-D scene completion","Edges as dense F-TSDF lift 3D semantic IoU by 6.9%"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the previously published scores for SSCNet, VVNetR-120, TNetFuse and the other baselines were computed under exactly the same test split, ground-truth voxelisation, and evaluation script as EdgeNet; the paper extracts those numbers from the original papers rather than running a shared evaluation.","fun_headline_variants_meta":{"raw":{"variants":["Edge binarization turns sparse RGB into dense 3D cues","F-TSDF edge volumes densify RGB for 3D scene completion","EdgeNet fuses depth and edge F-TSDF volumes for semantics","Binarized edges solve colour sparsity in RGB-D scene completion","Edges as dense F-TSDF lift 3D semantic IoU by 6.9%"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000575,"raw_usage":{"total_tokens":2664,"prompt_tokens":845,"completion_tokens":1819,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":461,"completion_tokens_details":{"reasoning_tokens":1699}},"tokens_in":461,"tokens_out":1819,"duration_ms":14486,"temperature":1.0,"reasoning_tokens":1699,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:30:51.526321+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run VVNetR-120 and TNetFuse with EdgeNet's re-rendered SUNCG test images and its NYU ground-truth voxelisation and evaluation code. If they reproduce their reported scores, EdgeNet's 70.3 and 33.7 IoU results hold; if their scores move, the claimed gains are protocol artefacts. A further check is to feed EdgeNet a random binary volume with the same density as the projected edges: if the IoU gain over EdgeNet-D persists, the improvement is not specifically due to RGB edges.","supporting_citations":[{"cited_title":"Semantic Scene Completion from a Single Depth Image,","cited_arxiv_id":null,"evidence_quote":"Defines the semantic scene completion task and the F-TSDF encoding that EdgeNet adopts for depth and edges."},{"cited_title":"A computational approach to edge detection,","cited_arxiv_id":null,"evidence_quote":"Provides the Canny edge detector used to turn RGB images into binary edge maps before 3D projection."},{"cited_title":"Semantic Scene Completion Combining Colour and Depth: preliminary experiments","cited_arxiv_id":"1802.04735","evidence_quote":"Shows that naive projection of raw RGB channels into the voxel grid fails to improve completion, motivating the edge encoding."},{"cited_title":"See and think: Disentangling semantic scene completion,","cited_arxiv_id":null,"evidence_quote":"Supplies TNetFuse and SNetFuse, the two-stage depth-plus-2D-segmentation approaches whose NYU results EdgeNet nearly matches."},{"cited_title":"View-V olume Network for Semantic Scene Completion from a Single Depth Image,","cited_arxiv_id":null,"evidence_quote":"Supplies VVNetR-120, the depth-only method that EdgeNet surpasses by 3.3 points on SUNCG."},{"cited_title":"U-Net: Convolutional net- works for biomedical image segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the U-Net encoder-decoder design on which EdgeNet's architecture is based."},{"cited_title":"Semantic scene completion with dense CRF from a single depth image,","cited_arxiv_id":null,"evidence_quote":"Supplies the DCRF depth-only baseline reported in the comparison tables."}],"review_version":1}