{"id":"b00c7813-1256-4b60-924d-3d41b3a85618","arxiv_id":"1908.05540","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":5,"one_line_summary":"A joint multi-task network performs monocular depth estimation and sparse depth completion with a shared two-stage architecture, reporting competitive-to-superior numbers on KITTI, but the evaluation omits a stronger baseline and may have train/test overlap.","lead":"This paper trains one network that both turns a single color photo into a distance map and fills in a sparse laser scan, reporting better results than several earlier methods on standard driving data. The headline numbers are credible only if the training and test images do not overlap, which the paper does not demonstrate.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 1 omits DORN, a stronger published baseline on the same Eigen split, so the 0.080 Abs Rel row cannot support the contemporary-SOTA claim; the unstated [54]/Eigen training overlap may further inflate the result.","rationale":"I read the paper as claiming that a two-stage joint network delivers state-of-the-art monocular depth estimation while remaining competitive at sparse depth completion. For that claim to hold, Table 1 must be a fair comparison against the strongest contemporary baselines and the training data must not include the test frames. The first condition is not met on the face of the paper: DORN [18] is a recognized SOTA method on the same KITTI Eigen benchmark and is cited in Related Work, yet it is absent from Table 1. Its published Abs Rel of 0.072 is lower than the paper's 0.080, so with a comparable protocol the headline superiority claim fails. The reader's identified concern about overlap between the [54] training sample and the Eigen test split is serious and should be checked, but it is secondary in emphasis because it requires file-list inspection, whereas the DORN omission is already visible in the paper itself. I therefore set agreement_with_reader to partial rather than agree. I do not move the verdict away from CONDITIONAL: the multi-task architecture and depth-completion numbers would still have value once the SOTA claim is reframed as competitive, but the conditions must explicitly include adding DORN to the comparison and enforcing a strict train/test separation.","tokens_in":13080,"tokens_out":10232,"duration_ms":107620,"concrete_test":"Run DORN's released model on the KITTI Eigen split using the same maximum-depth crop, ground-truth masking, and evaluation script used for Table 1, and record Abs Rel and RMSE. If DORN's numbers match its published 0.072 and 2.727, the paper's 0.080 Abs Rel row is not state of the art, and the central monocular superiority claim is refuted. As a secondary check in the same evaluation, count the file-level intersection between the training sample from [54] and the Eigen test split to test for leakage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing issue is the completeness of the monocular comparison. Section 4.1 presents Table 1 as evidence that the approach outperforms contemporary state-of-the-art depth estimation, but Table 1 does not include DORN [18], which the paper itself cites in Section 2.1 and which reports KITTI Eigen-split results of Abs Rel 0.072, Sq Rel 0.307, RMSE 2.727, RMSE log 0.120, and delta1 0.932. Against the paper's Our Approach row (Abs Rel 0.080, RMSE 4.437, delta1 0.929), DORN is better on every published metric. If those numbers are protocol-comparable, the central claim that the model outperforms contemporary state-of-the-art is false on its face. The contribution bullets restrict the comparison to [5,7,14,20,31,36,62,66], but the abstract and Table 1 caption make a broader SOTA claim. A second, compounding risk is data provenance: Section 3 says only that training images are randomly sampled from the depth-completion dataset [54], while Section 4.1 evaluates on the Eigen split from [14]; the text never states that Eigen test frames were excluded from the [54] sampling pool. Because [54] is assembled from KITTI raw sequences, at least partial overlap is likely, which would directly inflate the reported 0.080 Abs Rel. Section 5 also acknowledges degenerate upper-region content in outputs, so the quantitative advantage over prior work is not obviously robust to protocol changes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a two-stage deep network for jointly solving monocular depth estimation and sparse depth completion. A sparse generator (SG) first maps an RGB image to a sparse depth map, and a dense generator (DG) then maps that sparse map to a dense depth map; the two sub-networks are trained end to end on a mixture of synthetic data [17] and real KITTI depth-completion data [54], using reconstruction, adversarial, and smoothness losses. For monocular depth estimation, the full SG+DG chain is evaluated on the Eigen split of KITTI and reported to achieve Abs Rel 0.080 (Table 1). For depth completion, DG alone is evaluated on the KITTI depth-completion validation set, reporting RMSE 892 mm (Table 2). The paper also presents ablations and qualitative generalisation results on local urban data.","tokens_in":13430,"tokens_out":3294,"duration_ms":31311,"significance":"If the reported monocular results are protocol-clean, the architecture is a useful contribution: the idea of explicitly generating a sparse intermediate depth representation as a scaffold for dense prediction is simple and well-motivated, and the paper provides a reasonably thorough ablation (loss components, single vs two-stage, synthetic vs real training). The adversarial-training component and the mixed-domain training strategy are consistent with prior practice, and the quantitative comparison against several published methods on both tasks gives the reader a meaningful baseline. However, the headline claim of outperforming contemporary state of the art on monocular depth estimation is not yet established because the comparison omits a stronger published method and because the training/evaluation protocol does not rule out train-test overlap. The depth-completion results are competitive but not state of the art, so the abstract's phrasing 'superior performance across both problem domains' overstates the findings. No code or trained models are released, which limits reproducibility.","major_comments":[{"comment":"The claim that the approach outperforms 'contemporary state-of-the-art techniques' on monocular depth estimation is not supported by the evidence in Table 1 because DORN (Fu et al. [18], cited in the paper) is omitted. DORN reports Abs Rel 0.072, Sq Rel 0.307, RMSE 2.727, RMSE log 0.120, and delta1 0.932 on the same Eigen split, which is better than the paper's Our Approach row (Abs Rel 0.080, Sq Rel 0.836, RMSE 4.437, RMSE log 0.157, delta1 0.929) on every metric. If those numbers are protocol-comparable, the central monocular claim is false as stated. The authors should include DORN in the table, or, if there is a protocol difference (e.g., different crop or depth cap), state it explicitly and justify why the comparison is still fair.","section":"Section 4.1, Table 1"},{"comment":"The paper never states that the Eigen-split test images used in Table 1 were excluded from the training pool sampled from the KITTI depth-completion dataset [54]. The depth-completion dataset is assembled from KITTI raw sequences, and the Eigen test split is drawn from the same raw sequences, so at least partial overlap is likely if no exclusion was performed. This would directly inflate the reported monocular results. The authors must specify the exact exclusion procedure and, ideally, verify that no Eigen test frame appears in the training set, and then re-report Table 1 if any overlap is found.","section":"Section 3 (data sampling) and Section 4.1 (evaluation)"},{"comment":"The abstract claims the approach is capable of 'superior performance' 'across both problem domains,' but Table 2 shows that on the KITTI depth-completion validation set the method (RMSE 892 mm) ranks behind Van Gansbeke et al. [55] (802 mm) and Ma et al. [39] (879 mm). The contribution bullet in Section 1 is more careful in restricting the superiority claim to methods [10, 16, 40, 50, 54], but the abstract and the phrase 'outperforms contemporary state-of-the-art techniques across both problem domains' overstate the completion results. The authors should soften the wording to match the actual comparison, e.g., 'competitive with state-of-the-art' for completion.","section":"Section 4.2, Table 2 and Abstract"}],"minor_comments":[{"comment":"Typo: 'In order words' should be 'In other words'.","section":"Section 1"},{"comment":"The row 'Train Set Mean [19]' is not a published method but a trivial baseline; consider labeling it explicitly as a lower-bound baseline rather than a competing method.","section":"Section 4.1, Table 1"},{"comment":"In the SN/L1/Adv row, the accuracy metric column labeled shift <1.25^2 reports 0.862, which is lower than the preceding shift <1.25 value of 0.913; this is almost certainly a typographical error and should be corrected.","section":"Section 4.3, Table 3"},{"comment":"The grid-search procedure for the loss weights is only described as 'a basic grid search' without giving the ranges or step sizes; for reproducibility, provide the search space.","section":"Section 3.3"},{"comment":"No mention is made of code or model release. Given that the exact training/test split and the random sampling from [54] are central to the validity of the results, a code release or a precise data-split specification is strongly recommended.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The two main concerns are (i) the omission of DORN from the monocular comparison and (ii) the unstated possibility of train/test overlap between the depth-completion training pool and the Eigen-split test set. Both are addressable in revision, provided the authors are willing to update or re-run the experiments. The paper itself is otherwise methodologically straightforward; the multi-task idea is not deeply novel but is sensible. If the authors can demonstrate no overlap and can either add DORN or transparently explain why DORN is excluded (e.g., different depth range or crop), the monocular claim could be accepted as a competitive result rather than a state-of-the-art one. The depth-completion section is honest in the body but the abstract overstates it."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nMy read on 1908.05540: the two-stage design is genuinely useful. Chaining a sparse generator (RGB to sparse depth) into a dense generator (sparse to dense), trained end-to-end on a mix of KITTI depth completion and synthetic data, is a new combination of known blocks. The ablation shows the two-stage structure helps over a single network, and the depth completion results are honestly framed as competitive rather than SOTA. That part is solid.\n\nThe problems are in the monocular comparison. Table 1 claims to outperform contemporary state-of-the-art on the Eigen split, but it omits DORN, which the paper itself cites and which reports Abs Rel 0.072, RMSE 2.727, delta1 0.932 – better than the paper's 0.080, 4.437, 0.929 on every metric. If those numbers are protocol-comparable, the central claim is false as stated. The contribution bullet narrows the comparison list, but the abstract and table caption make a broader SOTA claim.\n\nThe second issue is data provenance. Section 3 says training images are randomly sampled from the depth-completion dataset [54]. The Eigen test split comes from the same KITTI raw sequences. The paper never states that Eigen test frames were excluded from the sampling pool. Without that statement, the reported 0.080 Abs Rel may be inflated by training on the test set. That is load-bearing, not a nit. Section 5 also admits degenerate upper-region content, so the quantitative edge is not obviously robust.\n\nWhat's genuinely good: the masked reconstruction loss for missing ground truth is sensible, the architecture is clearly described, and the ablation study gives a real sense of what matters. No code, no error bars – that limits reproducibility but is not a fatal flaw on its own.\n\nBottom line: this deserves a serious referee, but with major revisions. The two-stage idea is worth exploring, and the depth completion side is credible. The monocular SOTA claim needs re-evaluation with DORN included and a strict statement about train/test separation. I'd send it to review, flagging those two items as mandatory. I would not cite it until the protocol questions are resolved. Reading group: maybe, as a case study in how easy it is to publish an inflated SOTA claim.\n\nRecommendation: peer review, conditional on addressing the missing baseline and data provenance.\n\nBest,\n[Your name]","headline":"Two-stage joint depth network is a real idea, but the monocular SOTA claim is not supported once DORN and the train/test overlap question are considered.","tokens_in":13916,"tokens_out":2704,"would_cite":false,"duration_ms":26467,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T45","68T07"],"pacs":[],"model":"deepseek-v4-flash","headline":"A joint two-stage model that first hallucinates sparse LiDAR-style depth and then densifies it outperforms previous monocular depth estimators on KITTI while remaining competitive at LiDAR depth completion.","keywords":["monocular depth estimation","sparse depth completion","multi-task learning","generative adversarial networks","KITTI dataset","synthetic data","skip connections","autonomous driving"],"falsifier":"Retrain the identical model on the KITTI depth-completion training split with every image belonging to the monocular test split of [14] excluded, then evaluate on that test split; if Abs Rel rises well above the reported 0.080, the headline gain was inflated by training on the test distribution. A cheaper check is to compare the sequence identifiers of the sampled training frames against the test split.","tokens_in":12871,"feed_emoji":"🚗","tokens_out":6971,"duration_ms":66174,"temperature":0.7,"pith_summary":"Monocular depth estimation and sparse depth completion are usually treated as separate problems with separate networks. This paper claims they are two stages of one representation: a model that first hallucinates a LiDAR-style sparse depth map from an RGB image, then densifies that map, can outperform contemporary single-task methods on the standard monocular test split (Abs Rel 0.080) while the densifying stage alone stays competitive at completing real LiDAR depth (RMSE 892 mm). The authors argue that the intermediate sparse-to-dense task forces the model to learn both low-level structure and high-level scene context, and that mixing synthetic dense depth with real sparse ground truth supplies supervision at a density no real dataset alone provides. If correct, this matters because a single trained model could serve both camera-only and LiDAR-equipped platforms, and because an auxiliary completion objective appears to improve the primary estimation task rather than compete with it.","feed_headline":"Two-stage depth model tops prior monocular depth on KITTI","feed_subtitle":"RGB-only estimates beat earlier methods while the same network's dense half also completes LiDAR scans.","key_machinery":"The load-bearing object is the two-stage sparse-to-dense generator $G(x) = \\mathrm{DG}(\\mathrm{SG}(x))$: the sparse generator is an encoder-decoder with skip connections that emits a sparse depth map, and the dense generator is a residual-block encoder-decoder that consumes that sparse map and emits dense depth. The key identity is the composite loss $$\\mathcal{L} = \\lambda_{\\mathrm{recSG}}\\|\\mathrm{SG}(x)-y_s\\|_1 + \\lambda_{\\mathrm{recDG}}\\|M \\odot \\mathrm{DG}(\\mathrm{SG}(x))-y_d\\|_1 + \\lambda_{\\mathrm{adv}}\\mathcal{L}_{\\mathrm{adv}} + \\lambda_s \\mathcal{L}_s,$$ with the binary mask $M$ applied only to real-world dense ground truth. The mask lets incomplete real data supervise the model without penalizing missing sky and distant regions, while synthetic dense depth supplies the full-density signal that real data cannot provide.","core_discovery":"The central discovery is that a single end-to-end model can perform both tasks, and that the completion stage is not a burden but the source of the gain. Concretely, the generator $G(x) = \\mathrm{DG}(\\mathrm{SG}(x))$ translates RGB to sparse depth and then to dense depth; the sparse generator is trained with an $\\ell^1$ reconstruction loss against projected LiDAR-like sparse depth, while the dense generator is trained with a masked $\\ell^1$ loss against real dense ground truth (masking out missing pixels) and an unmasked loss against synthetic dense depth. Two discriminators, one per data domain, provide adversarial supervision, and an edge-aware smoothness term penalizes large depth gradients where image gradients are small. The paper reports that this full configuration reaches Abs Rel 0.080 on the standard monocular test split of [14], beating the listed contemporary methods, and RMSE 892 mm on the depth-completion validation split of [54], competitive with the listed completion-specific methods.","pith_inferences":["The paper does not state whether the sampled KITTI depth-completion training frames are disjoint from the standard monocular test images of [14]; if they are not, retraining on a strictly disjoint split is the decisive follow-up experiment.","If the sparse intermediate task is what carries the gain, then attaching an SG-like auxiliary head should improve other dense depth backbones, not just this architecture; the paper does not test that.","The dual-discriminator design suggests mixed-domain training is unstable under a single adversarial loss; a domain-weighting or adaptive discriminator might achieve the same benefit with less machinery.","Because the method only needs dense synthetic depth and sparse real depth, it may transfer to indoor RGB-D scenarios where dense ground truth is also incomplete; this is an untested extrapolation."],"forward_implications":["A camera-only vehicle could run the full chain for monocular depth, while a LiDAR-equipped vehicle could run only the dense generator, reusing one training run for both deployment modes.","Adding a sparse-completion auxiliary task is a viable recipe for boosting dense depth regression, even when the final output is dense depth from RGB alone.","Blending synthetic dense depth with real sparse LiDAR depth can substitute for dense real-world ground truth, which is rarely available at scale.","The reported runtimes (33.4 ms for monocular, 18.1 ms for completion) indicate the two-stage model operates near real-time on a single GPU, so the multi-task design is not purely academic."],"supporting_citations":[{"why":"Supplies the real-world training images with relatively dense but incomplete ground truth, defines the depth-completion validation set, and is the main completion baseline.","marker":"[54]"},{"why":"Supplies synthetic urban images with fully dense depth, enabling supervision where real ground truth is missing.","marker":"[17]"},{"why":"Supplies the KITTI real-world imagery used for monocular evaluation and the underlying sequences used in the test split.","marker":"[19]"},{"why":"Defines the standard monocular train/test split and provides the earliest deep multi-scale baseline against which all comparators are measured.","marker":"[14]"},{"why":"Provides the adversarial training formulation used to keep outputs sharp.","marker":"[21]"},{"why":"Is a prior monocular approach using style-transfer domain adaptation that this work extends in the mixed synthetic/real training design and a key numeric comparator.","marker":"[5]"},{"why":"Is an unsupervised monocular baseline with left-right consistency that the paper compares against.","marker":"[20]"}],"fun_headline_variants":["Multi-task depth net beats singles on KITTI","One model, two depth tasks: completion boosts estimates","Depth completion powers monocular depth accuracy","End-to-end depth duo outperforms prior on KITTI","Shared latent depth model tops KITTI benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central numerical claim assumes the training frames sampled from the KITTI depth-completion set are disjoint from the standard monocular test images of [14]; the paper never states this exclusion, and the completion set contains many of the same sequences, so the reported 0.080 Abs Rel could partly reflect training on the test data.","fun_headline_variants_meta":{"raw":{"variants":["Multi-task depth net beats singles on KITTI","One model, two depth tasks: completion boosts estimates","Depth completion powers monocular depth accuracy","End-to-end depth duo outperforms prior on KITTI","Shared latent depth model tops KITTI benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1382,"prompt_tokens":993,"completion_tokens":389,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":609,"completion_tokens_details":{"reasoning_tokens":315}},"tokens_in":609,"tokens_out":389,"duration_ms":4044,"temperature":1.0,"reasoning_tokens":315,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:09:33.890077+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the identical model on the KITTI depth-completion training split with every image belonging to the monocular test split of [14] excluded, then evaluate on that test split; if Abs Rel rises well above the reported 0.080, the headline gain was inflated by training on the test distribution. A cheaper check is to compare the sequence identifiers of the sampled training frames against the test split.","supporting_citations":[{"cited_title":"Uhrig, N","cited_arxiv_id":null,"evidence_quote":"Supplies the real-world training images with relatively dense but incomplete ground truth, defines the depth-completion validation set, and is the main completion baseline."},{"cited_title":"Francis, K","cited_arxiv_id":null,"evidence_quote":"Supplies synthetic urban images with fully dense depth, enabling supervision where real ground truth is missing."},{"cited_title":"Geiger, P","cited_arxiv_id":null,"evidence_quote":"Supplies the KITTI real-world imagery used for monocular evaluation and the underlying sequences used in the test split."},{"cited_title":"Eigen, C","cited_arxiv_id":null,"evidence_quote":"Defines the standard monocular train/test split and provides the earliest deep multi-scale baseline against which all comparators are measured."},{"cited_title":"Goodfellow, J","cited_arxiv_id":null,"evidence_quote":"Provides the adversarial training formulation used to keep outputs sharp."},{"cited_title":"Atapour-Abarghouei and T","cited_arxiv_id":null,"evidence_quote":"Is a prior monocular approach using style-transfer domain adaptation that this work extends in the mixed synthetic/real training design and a key numeric comparator."},{"cited_title":"Godard, O","cited_arxiv_id":null,"evidence_quote":"Is an unsupervised monocular baseline with left-right consistency that the paper compares against."}],"review_version":1}