{"id":"6fc6c128-101f-41f1-aadd-dbf87f431f08","arxiv_id":"2501.01049","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":3,"one_line_summary":"TS-SatMVSNet incorporates slope-based modules into a cascade MVS network, achieving small improvements in terrain height estimation but with a flawed loss formulation and selective evaluation.","lead":"A satellite stereo network that uses local height differences ('slope') to guide height hypothesis sampling and smoothing reports modest accuracy gains on two benchmarks. The core training signal, a slope direction loss, is described in a way that appears non-differentiable, and the claimed state-of-the-art results are not consistently supported by the reported tables.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The slope direction loss (Eq. 9) is computed through the non-differentiable argmax in Algorithm 1, so it cannot backpropagate to the network; without a described straight-through or soft relaxation, this stated contribution is a no-op.","rationale":"The reader's weakest_assumption identifies the same point I consider load-bearing: the slope direction loss cannot be optimized as described because it is built on a non-differentiable argmax. If the proposed gradient test confirms zero gradients, then the paper's stated end-to-end slope-direction supervision is a no-op, and the absence of this loss in Table 3 means the claimed third contribution is not empirically supported. The remaining components, slope-guided interval partition and height correction, show small improvements in Table 3, so the method is not wholly without value, but the central claim of an end-to-end slope-aware framework with three synergistic components is not supported as written. The paper also contains internally inconsistent SOTA statements, for example Table 1 shows <2.5m of 77.92% versus 77.93% for SatMVS(RED-Net), while the introduction claims at least a 5% improvement on this metric, and this compounds the correctness risk. I would not adjust the reader's REJECT verdict; at minimum, the paper would need a corrected, differentiable formulation of Eq. 9 and a proper ablation before its central claim could be accepted.","tokens_in":18990,"tokens_out":7683,"duration_ms":78290,"concrete_test":"In a small PyTorch script, construct the Algorithm 1 slope direction map from a 3x3 height tensor with requires_grad=True, compute Ls exactly as in Eq. 9, and call backward on Ls; inspect grad w.r.t. the height tensor. If the gradient is exactly zero (modulo floating-point tie effects), Eq. 9 cannot train the height network. As a secondary check, retrain the full model with and without Ls while keeping all other settings identical; since Table 3 already omits Ls, the authors should report this ablation and the resulting WHU-TLC metrics to substantiate any claimed benefit.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equation 9 defines an MSE loss between predicted and pseudo-GT slope direction maps, and the predicted map is produced by Algorithm 1 from the predicted height map. The only described route from heights to direction codes is an argmax over the nine 3x3 positions (indexmax <- Max(p3x3), then p3x3[indexmax]); indexing with that integer result has zero gradient w.r.t. the height values almost everywhere, with nonzero derivative only on tie sets of measure zero. The paper does not mention a straight-through estimator, soft argmax, Gumbel relaxation, or any other differentiable surrogate. Consequently Ls cannot provide a learning signal to the network through the slope direction branch, and Loverall (Eq. 10) reduces to the height loss plus a term whose gradient is zero. This is not merely a missing experiment: the ablation in Table 3 omits the slope direction loss entirely, so there is no evidence that it improves height accuracy, even though the method description presents it as a central component. The additional semantic issue is that direction codes 0-8 are categorical labels, so an L2 loss on them is not a principled objective even if a relaxation were added. Either the loss is non-differentiable as described, or the algorithm does not actually produce the claimed direction codes; both readings undermine the stated contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes TS-SatMVSNet, a coarse-to-fine multi-stage satellite multi-view stereo network for terrain height estimation. The core idea is to compute a slope map and a slope direction map from the predicted height map using a 3x3 maximum-height-difference heuristic, then use these in two modules: a slope-guided interval partition module that reallocates height hypothesis planes and a height correction module based on a learnable Gaussian filter. The paper also introduces a slope direction loss between predicted and pseudo-ground-truth slope direction maps. Experiments are reported on WHU-TLC and MVS3D, with quantitative comparisons against traditional and learning-based methods, ablations of the proposed modules, and qualitative visualizations.","tokens_in":19294,"tokens_out":6071,"duration_ms":57966,"significance":"The general idea of injecting terrain-specific inductive bias into satellite MVS height estimation is timely and potentially useful for large-scale Earth surface reconstruction, and the paper provides a reasonably broad empirical study spanning two datasets and multiple comparison methods. The ablation of the slope-guided interval partition and height correction modules shows measurable improvements over the baseline, which is a useful empirical result. However, the central technical claim is undermined by the non-differentiability of the slope direction loss as described, and several quantitative claims about state-of-the-art performance are not supported by the reported numbers, so the paper in its current form does not establish the claimed contributions.","major_comments":[{"comment":"The slope direction loss in Eq. (9) is not a valid training signal as described. Algorithm 1 computes the direction code by selecting the index of the maximum height in each 3x3 patch and then indexing p3x3 with that integer; this argmax/index operation has zero gradient almost everywhere with respect to the predicted height map. No straight-through estimator, soft-argmax, or other relaxation is mentioned. Consequently the gradient of L_s with respect to the network parameters is zero almost everywhere, and L_overall in Eq. (10) effectively reduces to the height loss. The claim that the slope direction loss 'implicitly optimizes height estimation' is unsupported, and Table 3 does not include an ablation row isolating this loss, so there is no experimental evidence for its contribution. Additionally, the direction codes 0-8 are categorical labels, so an L2 loss on them is semantically questionable even if a differentiable surrogate were introduced.","section":"Methodology, Algorithm 1 and Eq. (9)"},{"comment":"The state-of-the-art and quantitative improvement claims are not supported by the reported results. The Introduction states that the method exceeds other SatMVS-based methods by 'at least 16% in MAE metric and at least 5% in < 2.5m metric at WHU-TLC dataset,' but Table 1 shows the MAE improvement over SatMVS-F is only 1.895 to 1.879 (about 0.8%), and over SatMVS(RED-Net) it is 1.945 to 1.879 (about 3.4%). Moreover, on the <2.5m metric at 2048x1472, TS-SatMVSNet obtains 77.92% while SatMVS(RED-Net) obtains 77.93%, so the method is not strictly state-of-the-art on that metric. The abstract's unqualified 'state-of-the-art performance' claim should be revised to name specific metrics and baselines where the improvement is real.","section":"Introduction and Table 1"},{"comment":"The WHU-TLC* comparison in Table 3 does not support the claimed terrain-specific advantage because it changes the test set and the method simultaneously. The baseline rows are evaluated on WHU-TLC, while the final TS-SatMVSNet row is evaluated on WHU-TLC*, a subset constructed by the authors to contain only terrain areas. The reported improvements (MAE 2.026 to 1.793, <2.5m 77.01% to 79.03%) could reflect the easier, selected subset rather than the method. To support the claim, the same baselines must be evaluated on WHU-TLC* under identical conditions.","section":"Ablation Study, Table 3"}],"minor_comments":[{"comment":"Equation (7) uses the symbol ⊙ for what appears to be a convolution or sliding-window filtering operation, but the notation is not defined; also, the 'learnable Gaussian filter' has only a single learnable scalar i, so the description overstates the flexibility of this module.","section":"Methodology, Eq. (7)"},{"comment":"Algorithm 1 is underspecified: the branch S_diff == 0 is not handled, and the final assignment 'Abs(S_diff + p3x3[indexmax])' is unclear without additional pseudocode or a precise tensor-indexing explanation.","section":"Algorithm 1"},{"comment":"The Experimental Settings section contains an incomplete cross-reference ('in Sec .') and the MVS3D preprocessing pipeline is described only as a promise to open-source later; enough detail should be provided for reproducibility of the reported MVS3D results.","section":"Experimental Settings"},{"comment":"There are numerous typos and inconsistencies, including 'slop' for 'slope', 'Parition' for 'Partition', 'specifical' for 'specific', and duplicated table footnote lines in Table 2; a careful proofreading pass is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The non-differentiability of the slope direction loss is the most serious issue; if it cannot be demonstrated that the loss actually trains the network, the authors should remove or replace it and re-run the experiments. The state-of-the-art claim also needs to be reworded to match the numbers. The empirical evidence for the two slope-guided modules is suggestive, so the paper is not hopeless, but in its current form the central claims are not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the slope-guided interval partition is a small real improvement; the slope direction loss is broken as written and should not be counted as a contribution.\n\nWhat is actually new: the paper builds on SatMVS and adds three things. The slope-guided interval partition (SIPM) is the strongest. Allocating more height hypothesis planes to steeper sub-ranges is sensible, and Table 3 shows it moves MAE from 2.026 to 1.911 and <2.5m from 77.01 to 77.83 on WHU-TLC. That is a modest but consistent gain, and it survives in the full model. The height correction module (HCM) helps a little. The paper is honest about its urban-area limitation and the MVS3D results are competitive.\n\nThe load-bearing problem is the slope direction loss. Algorithm 1 as written does not produce directional codes. 'indexmax <- Max(p3x3)' appears to store the maximum value, not its index; that value is then used to index p3x3, and the conditional arithmetic sets the difference to 0 for any positive slope. The final output is Abs(Sdiff + max), which is a height value, not one of the nine codes in Figure 4. Even if we charitably read 'indexmax' as argmax, the operation is non-differentiable, and the paper describes no straight-through estimator, soft-argmax, or Gumbel relaxation. So Eq. 9 has no gradient path to the network; the slope direction loss is a no-op as described. The ablation omits this loss entirely, so there is no evidence it helps. L2 on categorical codes is also semantically arbitrary.\n\nThere are smaller issues. Table 1 shows <2.5m 77.92 vs 77.93 for SatMVS(RED-Net), so the abstract's 'state-of-the-art' claim is overstated (the paper acknowledges this, but only later). The WHU-TLC* sub-dataset is a selected subset of terrain-index folders, which risks selection bias, and the selection criteria are not given in detail. The claimed 16% MAE improvement in the introduction does not match Table 1.\n\nWho this is for: satellite MVS and terrain DSM researchers might borrow the SIPM idea. But the paper in its current form is not acceptable as a three-component framework; it needs a major revision to either remove the slope direction loss or specify a differentiable version and ablate it, and to scale back the SOTA claim. I would not desk-reject this—the SIPM result is real enough to deserve referee time—but a careful reviewer should push hard on the loss.","headline":"A useful slope-guided interval partition buried under a slope direction loss that cannot backpropagate as written.","tokens_in":19811,"tokens_out":5291,"would_cite":false,"duration_ms":49029,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that feeding slope information, computed from each stage's height map, back into a coarse-to-fine satellite MVS network improves terrain height estimation, reporting state-of-the-art MAE of 1.879 m on WHU-TLC and…","keywords":["satellite multi-view stereo","terrain height estimation","slope-aware MVS","RPC warping","digital surface model","height hypothesis intervals","Gaussian height correction","WHU-TLC dataset"],"falsifier":"To test the central mechanism, hold the pseudo ground-truth slope-direction map fixed and compute the gradient of the slope-direction loss with respect to the predicted height map. If the discrete max-and-code procedure is implemented as described, that gradient is zero almost everywhere, so a training run with the slope-direction loss removed should give identical heights to one with it enabled; if it does, the loss is inert and any gains come from the two slope-guided modules and the height loss.","tokens_in":18753,"feed_emoji":"🏔️","tokens_out":10370,"duration_ms":86714,"temperature":0.7,"pith_summary":"This paper argues that general-purpose learning-based multi-view stereo (MVS) pipelines for satellite imagery ignore a basic fact about the Earth's surface: terrain is gently undulating, and its local steepness can be measured by slope. The authors propose TS-SatMVSNet, an end-to-end three-stage MVS network that computes a slope map and a slope-direction map from its own predicted height map using a 3x3 max-absolute-difference heuristic, then uses those maps to reallocate per-pixel height hypothesis planes, to correct heights with a learnable Gaussian filter, and to add a slope-direction loss against a pseudo ground truth derived from the GT height map. On WHU-TLC, the method reports MAE of 1.879 m against 1.895 m for SatMVS-F and raises the fraction of pixels within 7.5 m of ground truth from 80.05% to 97.34%; on MVS3D it reports the best average <1.0 m, Median, and RMSE among compared methods. If correct, the work would show that cheap terrain geometry priors like slope can be folded into existing MVS pipelines to reduce artifacts and improve DSM completeness without extra supervision data.","feed_headline":"Slope-aware stereo net tightens satellite terrain heights","feed_subtitle":"It routes a 3x3 slope heuristic through a multi-stage MVS net, reporting MAE of 1.879 m on WHU-TLC.","key_machinery":"The Height-based Slope Calculation Strategy (HSCS) carries the argument: for each pixel $x$ with 3x3 height patch $p_{3\\times 3}(x)$, it defines the slope as $S(x) = |\\max(p_{3\\times 3}(x)) - H(x)|$ and the slope direction as the integer 0-8 pointing from the center to the max-height cell, with code 4 meaning the center is the maximum. Because this map is computed from the network's own height predictions, no external slope labels are needed. The slope map feeds the Slope-guided Interval Partition Module, which splits each pixel's $[H-\\sigma, H+\\sigma]$ height range into lower and upper sub-ranges weighted by $S_{\\min}$ and $S_{\\max}$, placing more hypothesis planes on the steeper side. The Height Correction Module applies a 3x3 learnable Gaussian kernel, parameterized as a fixed 1/16, 1/8, 1/4 pattern scaled by a learned multiplier, to smooth abnormal height values. The slope-direction map is then compared to a pseudo ground-truth slope-direction map, computed from the GT height map, through an L2 loss.","core_discovery":"The central claim is that integrating slope information into a satellite MVS pipeline materially improves terrain height estimation. The paper's specific discovery is that a height map itself can be turned into a slope representation with no additional supervision: for each pixel, the slope is the absolute difference between the center height and the maximum of its eight neighbors, and the slope direction is the direction from the center toward that maximum, encoded as one of nine integer codes. That representation drives two novel modules, a slope-guided interval partition and a learnable Gaussian height correction, plus a slope-direction loss, allowing the coarse-to-fine stages to adapt to terrain undulation. The authors summarize the contribution as a complete slope-aware height estimation network that reports state-of-the-art accuracy on WHU-TLC and competitive generalization on MVS3D.","pith_inferences":["One extension the paper leaves implicit: the one-sided max-difference slope operator could be replaced by a symmetric gradient or Sobel-style estimate, which would treat upward and downward terrain changes identically and might soften the urban artifacts the paper reports.","A straight-through estimator or regression to a direction vector would give the slope-direction loss a meaningful gradient through the discrete direction-encoding step; without such a mechanism, the loss cannot be assumed to train the network.","The same two slope-guided modules could be grafted onto other MVS or dense-stereo backbones; the paper only tests them inside its three-stage RPC pipeline, so whether they transfer across architectures remains open.","Given the paper's own admission that urban scenes break the small-plane assumption, a learned gate that turns slope guidance off at abrupt edges is a natural next test."],"forward_implications":["The paper's ablations show that adding slope-guided interval partitioning to the baseline improves MAE from 2.026 m to 1.911 m on WHU-TLC; adding the height correction module alone improves MAE to 2.002 m.","On the full WHU-TLC test set the model reports MAE 1.879 m and RMSE 3.892 m, versus 1.895 m and 3.654 m for SatMVS-F, and <7.5 m accuracy rises to 97.34% from 80.05%.","Accuracy stays nearly constant across resolutions: moving from 2048x1472 to 5120x5120 changes the paper's metrics by only 0.1-0.6%, whereas SatMVS(RED-Net) fluctuates by up to 15.9%.","On MVS3D the WHU-TLC-pretrained model reports the best average <1.0 m (60.635%), Median (0.353 m), and RMSE (2.898 m), indicating that slope-aware pretraining transfers to a different satellite sensor without fine-tuning.","Because the slope maps come from the network's own height estimates rather than external labels, the same mechanism supplies training supervision and inference-time refinement without additional annotation effort."],"supporting_citations":[{"why":"Supplies the differentiable RPC warping module and the WHU-TLC dataset that the entire pipeline is built on and evaluated against.","marker":"(Gao et al. 2021)"},{"why":"Provides the Sat-MVSF training/evaluation protocol and the primary comparison numbers for both WHU-TLC and MVS3D.","marker":"(Gao et al. 2023)"},{"why":"Provides the MVS3D benchmark and LiDAR ground truth used to test cross-sensor generalization.","marker":"(Bosch et al. 2016)"},{"why":"CasMVSNet is the cascade cost-volume backbone that SatMVS adapts and one of the deep baselines the paper compares against.","marker":"(Gu et al. 2020)"},{"why":"UCS-Net contributes the adaptive-thin-volume idea behind the pixel-wise height range that the slope-guided partition modifies.","marker":"(Cheng et al. 2020)"},{"why":"Defines the all-pixel and pixel-wise height range formulas reused in Eqs. 2-3 that the slope-guided interval partition reallocates.","marker":"(Zhang et al. 2023)"},{"why":"RED-Net(RPC) is a key deep-learning baseline whose resolution sensitivity is compared against the proposed model's stability.","marker":"(Liu and Ji 2020)"}],"fun_headline_variants":["Slope maps sharpen satellite terrain 3D reconstruction","Terrain slopes guide MVS for sharper satellite heights","Using terrain slope to boost MVS height accuracy","Height-derived slopes drive precise terrain reconstruction","Slope-aware MVS net tops WHU-TLC height accuracy"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the slope-direction loss can train the network: gradients must flow from that loss back through the discrete step that picks the tallest neighbor and assigns its direction code, and that step as written gives the predicted heights no gradient.","fun_headline_variants_meta":{"raw":{"variants":["Slope maps sharpen satellite terrain 3D reconstruction","Terrain slopes guide MVS for sharper satellite heights","Using terrain slope to boost MVS height accuracy","Height-derived slopes drive precise terrain reconstruction","Slope-aware MVS net tops WHU-TLC height accuracy"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000224,"raw_usage":{"total_tokens":1494,"prompt_tokens":1009,"completion_tokens":485,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":625,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":625,"tokens_out":485,"duration_ms":4849,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:35:24.434703+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"To test the central mechanism, hold the pseudo ground-truth slope-direction map fixed and compute the gradient of the slope-direction loss with respect to the predicted height map. If the discrete max-and-code procedure is implemented as described, that gradient is zero almost everywhere, so a training run with the slope-direction loss removed should give identical heights to one with it enabled; if it does, the loss is inert and any gains come from the two slope-guided modules and the height loss.","supporting_citations":[{"cited_title":", author Liu, J","cited_arxiv_id":null,"evidence_quote":"Supplies the differentiable RPC warping module and the WHU-TLC dataset that the entire pipeline is built on and evaluated against."},{"cited_title":", author Liu, J","cited_arxiv_id":null,"evidence_quote":"Provides the Sat-MVSF training/evaluation protocol and the primary comparison numbers for both WHU-TLC and MVS3D."},{"cited_title":", author Kurtz, Z","cited_arxiv_id":null,"evidence_quote":"Provides the MVS3D benchmark and LiDAR ground truth used to test cross-sensor generalization."},{"cited_title":", author Xu, Z","cited_arxiv_id":null,"evidence_quote":"UCS-Net contributes the adaptive-thin-volume idea behind the pixel-wise height range that the slope-guided partition modifies."},{"cited_title":", author Xu, W","cited_arxiv_id":null,"evidence_quote":"Defines the all-pixel and pixel-wise height range formulas reused in Eqs. 2-3 that the slope-guided interval partition reallocates."},{"cited_title":", author Ji, S","cited_arxiv_id":null,"evidence_quote":"RED-Net(RPC) is a key deep-learning baseline whose resolution sensitivity is compared against the proposed model's stability."}],"review_version":1}