{"id":"f78d661f-10b5-48c3-ac3f-67aeb1ba0a67","arxiv_id":"1908.09346","paper_version":4,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Dedge-AGMNet combines a depth edge auxiliary branch with an atrous granular multi-scale 3D aggregation module, yet its stated state-of-the-art claim conflicts with its reported fourth-place KITTI rankings.","lead":"A stereo matching network uses depth edge cues from semantic and instance labels to improve disparity estimates in difficult regions, reporting gains on standard benchmarks. The paper's own results contradict its headline claim of state-of-the-art performance, and a key parameter-count derivation is incorrect.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Section 4.5 contradicts the Abstract: Dedge-AGMNet is reported to rank fourth on both KITTI 2012 and KITTI 2015, so the central 'state-of-the-art' claim is internally refuted.","rationale":"I read the paper as claiming a new multi-benchmark state-of-the-art stereo matcher whose AGM module is parameter-cheap. The strongest support is internal: Table 6 shows the lowest SceneFlow EPE among listed methods, and the ablation tables show monotone gains when modules are added. Those are real evidence for the components. The KITTI claim, however, is contradicted within the same manuscript: Section 4.5 states the KITTI 2012 and KITTI 2015 test results both rank fourth, while the Abstract and conclusion call them state-of-the-art. This is not a matter of external consensus or missing comparison; the paper's own rank excludes SOTA. I therefore cannot accept the central claim. I do not rest the rejection on Eq. 5, as the reader's weakest-assumption field does: the formula is carelessly written and the exact parameter ratio depends on whether the first granular group has its own convolution, but under a charitable reading the granular convolution still costs less than standard convolution, so the efficiency claim is at most imprecisely quantified rather than refuted by the algebra alone. I also note Table 2's best KITTI validation uses a=0, so the auxiliary depth-edge loss is disabled in the best configuration; this weakens the causal story but is secondary to the false SOTA claim. Overall I agree with the reader's REJECT verdict, though for a different primary reason.","tokens_in":10428,"tokens_out":12594,"duration_ms":126300,"concrete_test":"Check the paper's own Section 4.5 and the KITTI leaderboards current at the v4 submission (March 2020): count the methods with KITTI 2015 D1-all below 1.85% and KITTI 2012 Out-Noc below 1.26%. If at least three such methods exist on each leaderboard, the Abstract's multi-benchmark state-of-the-art claim is false exactly as the text's 'rank fourth' admits.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing claim is the Abstract's assertion that Dedge-AGMNet 'outperforms other stereo matching networks and advances state-of-the-art performances on the SceneFlow, KITTI 2012 and KITTI 2015 benchmark datasets.' The paper's own results refute the KITTI part. Section 4.5 says the KITTI 2015 D1-all of 1.85% 'ranks in the fourth place' and the KITTI 2012 result 'ranks fourth, too' (Tables 5 and 4). Fourth place on each leaderboard is not state-of-the-art; at least three published methods have lower error. The SceneFlow first-place claim is supported by Table 6, but the central multi-benchmark SOTA claim is not. A secondary internal contradiction strengthens the concern: in Table 2, the best fine-tuned KITTI validation result is obtained at a=0, i.e. with the depth-edge loss term disabled, undercutting the causal claim that the depth-edge auxiliary task drives the reported gains. Additionally, Eq. 5's parameter-saving derivation is not rigorous, but the headline SOTA failure is decisive on its own.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Dedge-AGMNet, an end-to-end stereo matching network with two main contributions: a depth-edge auxiliary branch that shares features with the disparity branch and feeds them into a modified Dedge-SPP, and an atrous granular multi-scale (AGM) module built from granular convolutions intended to capture multi-scale context with fewer parameters than standard 3D convolutions. The authors generate depth-edge labels from semantic and instance segmentation maps, train on SceneFlow, Cityscapes, and KITTI, and report results on SceneFlow and the KITTI stereo benchmarks. The abstract and conclusion claim state-of-the-art performance on SceneFlow, KITTI 2012, and KITTI 2015, while the body reports first place on SceneFlow and fourth place on both KITTI leaderboards.","tokens_in":10686,"tokens_out":4922,"duration_ms":47418,"significance":"If the claimed contributions held, the paper would offer a practical multi-task architecture for stereo matching with a cheaper multi-scale aggregator and a principled depth-edge auxiliary task. The paper does provide useful experimental material: ablation studies on the depth-edge branch, the Dedge-SPP module, the AGM dilation configuration, and comparisons on public benchmarks. These ablations and the SceneFlow first-place result are concrete strengths. However, the central state-of-the-art claim is contradicted by the paper's own KITTI leaderboard results, and the parameter-count derivation that motivates the AGM module is algebraically incorrect. Because these two issues affect the paper's main claims, the contribution is not established as presented.","major_comments":[{"comment":"The abstract and conclusion state that Dedge-AGMNet 'outperforms other stereo matching networks and advances state-of-the-art performances on the Sceneflow, KITTI 2012 and KITTI 2015 benchmark datasets,' but Section 4.5 reports that the KITTI 2015 D1-all of 1.85% 'ranks in the fourth place' and that the KITTI 2012 result 'ranks fourth, too,' as shown in Tables 5 and 4. Fourth place on a public leaderboard is not state-of-the-art; at least three published methods have lower error on each benchmark. The multi-benchmark SOTA claim is therefore internally refuted and must be corrected, but the correction would remove the paper's headline contribution.","section":"Abstract and Section 4.5"},{"comment":"The derivation that granular convolution has approximately 1/G times the parameters of standard convolution is algebraically wrong. With Cin=Cout=C and kernel size s, the cascade of G group convolutions, each with weights in R^{(C/G)×(C/G)×s×s}, has G·(C/G)^2·s^2 = C^2 s^2 / G parameters before the pointwise convolution, plus C^2 for the pointwise term. The expression in Eq. (5) instead contains a factor (G−1)/G^2 and treats the pointwise term as negligible; neither the factor nor the neglect is justified. For C=32, s=3, G=4, the actual ratio to a standard convolution is (9/4 + 1)/9 ≈ 0.36, not 1/4. Since the AGM module's design rationale and the paper's efficiency claim rest on this equation, the claim that granular convolution captures multi-scale information with fewer parameters is not supported.","section":"Eq. (5), Section 3.3.2"},{"comment":"The paper attributes part of the accuracy gain to the depth-edge auxiliary task, but Table 2 reports that on the KITTI 2015 validation set the best D1-all is obtained at a=0, which disables the depth-edge loss term Ledge in Eq. (8). Under that setting the auxiliary task has no direct loss contribution, so the experiment does not demonstrate that the depth-edge loss drives the reported fine-tuned improvement. This weakens a load-bearing causal claim of the paper and needs to be addressed explicitly, either by reporting results for a=0 in the ablation table or by explaining which other mechanism (e.g., Dedge-SPP) carries the benefit.","section":"Table 2 and Section 4.3"}],"minor_comments":[{"comment":"The running-latency relation 'RLparallel = G−1 = 1/K × Rcascade' is dimensionally inconsistent: the left side is presented as a ratio or count, while the right side includes 1/K times a latency; please clarify the definitions of K and the intended units.","section":"Section 3.3.3"},{"comment":"In the row for EdgeStereo[17], the Non-Occluded D1-all value is reported as 3.00 while the All D1-all is 2.16; this is physically impossible, since non-occluded error cannot exceed the all-pixel error. This likely is a typographical error but should be corrected.","section":"Table 5"},{"comment":"The dataset name is written inconsistently as 'Sceneflow', 'Sceneflow', and 'SceneFlow'; please use one spelling consistently.","section":"Throughout"},{"comment":"The shared-concatenation notation '{F5(1),F,F5(2),F,...,F5(K),F}' is difficult to parse; a formal definition of the concatenation order would improve clarity.","section":"Section 3.2.2"}],"recommendation":"reject","confidential_remarks":"The paper's central claim of state-of-the-art performance is contradicted by its own leaderboard results, and the parameter-efficiency argument in Eq. (5) is not mathematically sound. These are load-bearing issues rather than presentation problems. Even if the writing were improved, the headline contribution would no longer be 'state-of-the-art' on KITTI 2012/2015, and the efficiency claim would need a substantially reworked derivation. I therefore recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The abstract promises state-of-the-art on SceneFlow, KITTI 2012, and KITTI 2015; the paper's own Section 4.5 says it ranks fourth on both KITTI leaderboards. That is not a minor exaggeration—it is an internal refutation of the central claim. The SceneFlow number (0.520 EPE) does look genuinely strong, but the multi-benchmark SOTA claim is simply false.\n\nWhat is actually new is the depth edge label: the union of semantic and instance boundaries, used as auxiliary supervision and fused via Dedge-SPP. That is a reasonable idea, and the ablations (Table 1) show it helps on SceneFlow and slightly on KITTI validation. The AGM module—parallel 3D granular convolutions with dilation rates 1, 4, 8, 16—also appears to give a nice improvement over PSMNet in their setup. I credit these components.\n\nThe soft spots are load-bearing. First, the parameter-count derivation in Eq. 5 is algebraically wrong. For C_in = C_out = C, the formula they write gives roughly s^2(G-1) + C^2, not (1/G)C^2s^2. So the claim that granular convolution uses about 1/G the parameters is not established; the actual savings (if any) depend on the exact grouping and point-wise projection in a way the paper does not rigorously analyze. Second, Table 2 shows the best KITTI validation result at a=0, i.e., with the depth edge loss term disabled. That undercuts the causal story for the auxiliary loss, even if the shared feature extraction and Dedge-SPP still help. Third, there are presentation issues: Adam beta1=0.09 looks like a typo, and the writing in places is hard to follow.\n\nThe paper is not a waste of time—the depth edge construction and the multi-scale aggregation idea could be worth pursuing. But as written, the claims are overstated and the math does not check out. If this lands on my desk, I would reject it and point the authors to fix the SOTA framing, redo the parameter analysis carefully, and clarify the role of the loss weight. It might become a solid workshop or second-tier paper after that.","headline":"A genuinely interesting depth-edge auxiliary idea and good SceneFlow numbers, but the KITTI SOTA claim is contradicted by the paper's own leaderboard results and the granular-convolution parameter math is wrong.","tokens_in":11227,"tokens_out":5356,"would_cite":false,"duration_ms":49529,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Depth-edge supervision plus granular 3D aggregation pushes stereo matching accuracy to reported state-of-the-art levels on SceneFlow and KITTI.","keywords":["stereo matching","disparity estimation","depth edge","multi-task learning","granular convolution","atrous convolution","multi-scale aggregation","cost volume"],"falsifier":"Evaluate Eq. (5) at the channel and kernel dimensions the deployed AGM actually uses (for example C=32, kernel size 3, G=4) and count the trainable weights of one granular 3D convolution directly; if the measured ratio to a standard 3D convolution is not about 1/G, the paper's efficiency claim is false. A behavioral check: replace AGM with a standard 3D convolution matched to the same parameter budget and rerun the SceneFlow ablation; equal-or-better EPE from the standard convolution would show the granular structure is not the source of the improvement.","tokens_in":10205,"feed_emoji":"👁️","tokens_out":12964,"duration_ms":123727,"temperature":0.7,"pith_summary":"This paper claims that stereo depth estimation improves when the network is guided by a depth edge auxiliary task and aggregates cost volumes with a granular multi-scale module. The authors create depth edge ground truth by combining semantic and instance boundaries, share the feature extractor between the disparity and edge branches, and inject edge features through a modified spatial pyramid pooling stage called Dedge-SPP. They also introduce the AGM module, which applies granular convolution in 3D with parallel atrous branches to widen the receptive field while using fewer parameters. In their experiments, the full network reports the top published end-point-error on SceneFlow and fourth place on the KITTI 2012 and KITTI 2015 stereo test sets. If these results hold, the recipe provides a concrete way to improve disparity estimation in ambiguous boundary regions without requiring an edge network at inference time.","feed_headline":"Depth-edge cues sharpen stereo matching on SceneFlow and KITTI","feed_subtitle":"A shared depth-edge branch plus granular 3D modules lowers disparity error; top SceneFlow EPE, fourth on KITTI.","key_machinery":"The load-bearing mechanisms are the depth edge auxiliary branch and the AGM module. Depth edges are binary maps generated by overlaying instance boundaries on semantic boundaries, so overlapping objects of the same class remain separated; the branch shares weights with the main feature extractor and contributes a binary cross-entropy loss plus a disparity-gradient consistency loss, and Dedge-SPP concatenates edge features into the spatial pyramid pooling stage. The AGM module is built from granular convolution, where input channels are split into G groups and each group's output is fed into the next group's filter, giving multiple effective receptive fields; in AGM this is extended to 3D convolution and run as four parallel atrous branches inside an hourglass-style block. The machinery's purpose is to supply geometric boundary cues that suppress non-depth edge noise while aggregating multi-scale cost-volume context with claimed savings in parameters and computation.","core_discovery":"The paper's central discovery, as the authors state it, is that depth edges—binary boundaries obtained by mining semantic and instance labels together—carry the supervision needed to resolve ill-posed stereo regions, and that an atrous granular multi-scale 3D module (AGM) is an effective and parameter-light aggregator. The depth edge branch shares its feature extractor with the disparity branch through hard parameter sharing, and Dedge-SPP embeds the edge features into the pyramid pooling stage. The AGM module repurposes granular convolution from 2D to 3D cost-volume aggregation, arranging four dilated granular branches in parallel to capture multi-scale context. The authors report that this combination reduces SceneFlow end-point-error by 27.0% relative to the pyramid stereo matching baseline, lowers KITTI 2015 validation D1-all from 1.67% to 1.54%, and places the method fourth on the KITTI 2012 and 2015 online test benchmarks while ranking first on SceneFlow among the compared published methods.","pith_inferences":["Editorial inference: the same depth-edge ground-truth recipe could be used to retrofit other cost-volume stereo networks, including real-time ones, because the edge branch is removed at inference and only the shared feature extractor and Dedge-SPP remain.","Editorial inference: if the parameter-saving formula is corrected and verified, a natural next test is to sweep the number of granular groups or atrous rates and measure the accuracy-latency frontier on embedded hardware; the paper fixes G=K=4 without such a sweep.","Editorial inference: the depth edge label could be generated from predicted semantic and instance masks rather than ground truth, enabling the auxiliary task for unlabeled or weakly labeled driving datasets; the paper only creates labels where semantic and instance ground truth exist.","Editorial inference: the loss-weight experiment on KITTI 2015 shows a=0 is best during fine-tuning, suggesting the depth-edge loss mainly helps during pre-training; a testable extension is to schedule the edge loss weight to decay during fine-tuning."],"forward_implications":["Depth edge supervision can be added to an existing stereo network at little extra inference cost, since the edge branch is active only during training and Dedge-SPP adds about 0.1M parameters.","The AGM module with four parallel atrous granular convolutions at dilation rates 1, 4, 8, and 16 outperforms both the hourglass baseline and other dilation combinations on SceneFlow end-point-error.","The full model reports a 27.0% SceneFlow EPE reduction and a 17.4% KITTI 2015 validation D1-all reduction relative to the pyramid stereo baseline.","On public online benchmarks, the method reports the top published SceneFlow EPE among the compared methods (0.520) and fourth place on both KITTI 2012 and KITTI 2015 test sets.","Pre-training on an additional urban driving dataset in combination with SceneFlow further lowers the KITTI 2015 validation error, suggesting driving-domain data reinforces the depth-edge cues."],"supporting_citations":[{"why":"Supplies the base ResNet-like feature extractor, spatial pyramid pooling, stacked hourglass 3D aggregation, and the main performance baseline the paper improves on.","marker":"[2]"},{"why":"Introduces granular convolution as a multi-scale residual block; the paper extracts and extends this idea to 3D for the AGM module.","marker":"[6]"},{"why":"The edge-supervised multi-task stereo network whose depth edge approach the paper adapts and compares against; motivates the depth edge branch.","marker":"[16]"},{"why":"Provides the shared-concatenation architecture that the depth edge subnetwork modifies for binary depth edge prediction.","marker":"[21]"},{"why":"Defines the concatenation cost volume and end-to-end regression used in the paper's cost volume and disparity prediction.","marker":"[10]"},{"why":"The atrous multi-scale 3D stereo network that motivates the parallel multi-scale AGM structure and serves as a comparison baseline.","marker":"[5]"},{"why":"Used in the embedding experiment to show the depth edge auxiliary branch transfers to another cost-volume stereo backbone.","marker":"[7]"},{"why":"Contributes the distance cost volume and dilated 3D hourglass ideas that the paper combines into its cost volume and AGM.","marker":"[1]"}],"fun_headline_variants":["Depth edges sharpen stereo disparity with AGM Net","Edge-aware stereo net tops SceneFlow, strong on KITTI","Dedge-AGMNet depth edges reduce stereo error","Multi-scale 3D aggregation with edge cues wins stereo","Shared edge branch boosts stereo matching accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the AGM module's granular convolution genuinely uses about 1/G the parameters of a standard convolution at the channel and kernel sizes used in practice; if Eq. (5)'s parameter count does not hold for the actual network dimensions, the efficiency rationale for the module collapses.","fun_headline_variants_meta":{"raw":{"variants":["Depth edges sharpen stereo disparity with AGM Net","Edge-aware stereo net tops SceneFlow, strong on KITTI","Dedge-AGMNet depth edges reduce stereo error","Multi-scale 3D aggregation with edge cues wins stereo","Shared edge branch boosts stereo matching accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000729,"raw_usage":{"total_tokens":3263,"prompt_tokens":944,"completion_tokens":2319,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2242}},"tokens_in":560,"tokens_out":2319,"duration_ms":17282,"temperature":1.0,"reasoning_tokens":2242,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T11:14:25.111092+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate Eq. (5) at the channel and kernel dimensions the deployed AGM actually uses (for example C=32, kernel size 3, G=4) and count the trainable weights of one granular 3D convolution directly; if the measured ratio to a standard 3D convolution is not about 1/G, the paper's efficiency claim is false. A behavioral check: replace AGM with a standard 3D convolution matched to the same parameter budget and rerun the SceneFlow ablation; equal-or-better EPE from the standard convolution would show the granular structure is not the source of the improvement.","supporting_citations":[{"cited_title":"5410–5418, (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the base ResNet-like feature extractor, spatial pyramid pooling, stacked hourglass 3D aggregation, and the main performance baseline the paper improves on."},{"cited_title":"EdgeStereo: An Effective Multi-Task Learning Network for Stereo Matching and Edge Detection","cited_arxiv_id":"1903.01700","evidence_quote":"The edge-supervised multi-task stereo network whose depth edge approach the paper adapts and compares against; motivates the depth edge branch."},{"cited_title":"5964–5973, (2017)","cited_arxiv_id":null,"evidence_quote":"Provides the shared-concatenation architecture that the depth edge subnetwork modifies for binary depth edge prediction."},{"cited_title":"66–75, (2017)","cited_arxiv_id":null,"evidence_quote":"Defines the concatenation cost volume and end-to-end regression used in the paper's cost volume and disparity prediction."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Used in the embedding experiment to show the depth edge auxiliary branch transfers to another cost-volume stereo backbone."},{"cited_title":"11786–11795, (2019)","cited_arxiv_id":null,"evidence_quote":"Contributes the distance cost volume and dilated 3D hourglass ideas that the paper combines into its cost volume and AGM."}],"review_version":1}