{"id":"d974197e-108a-4521-85f0-8f16268a9eed","arxiv_id":"1909.01028","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A monocular depth estimator that adds stereo left-right reconstruction and disparity consistency losses to an ego-motion-based view synthesis framework achieves modest gains on KITTI depth metrics.","lead":"This paper trains a neural network to estimate the depth of each pixel in a single driving video frame without using depth labels. It combines two known self-supervision signals, temporal video motion and stereo left-right image reconstruction, and reports improved accuracy on the KITTI benchmark.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '3/5 training data' claim is confounded: stereo right-image supervision and Cityscapes pretraining are not controlled in the Table 1 comparison, and the paper's own counts (16,384 vs 40,109) are 41%, not 60%.","rationale":"The reader's weakest assumption is exactly the load-bearing point. The paper's central efficiency claim is not merely weakened; it is not identifiable from the reported experiment, because the comparison to Zhou et al. varies at least three things at once: the proposed stereo losses, the right-image supervision signal, and Cityscapes pretraining (Section IV.A). The paper itself concedes the dependence on stereo pairs in the limitations. The internal inconsistency between the reported counts (16,384 / 40,109 = 0.41) and the headline fractions (3/5, half) is a separate, concrete red flag: the authors do not even agree on what the claim is. A single controlled ablation could test the data-efficiency claim, but until that is done, the abstract's '3/5' assertion is unsupported. This does not require questioning anyone's integrity; it is a straightforward missing-control problem. If the authors ran the proposed ablation and released the code, the contribution could be reconfigured as an incremental engineering result, but as submitted the central claim fails. I agree with the reader's REJECT and would not change that verdict.","tokens_in":11277,"tokens_out":6166,"duration_ms":55797,"concrete_test":"Run a controlled comparison on the KITTI Eigen split: train the baseline Zhou et al. method on the exact 16,384-sequence subset used in Section IV.A, with the same Cityscapes pretraining, epoch count, learning-rate schedule, and hyperparameters, but without any stereo losses, and evaluate on the same test split. If this baseline reaches Abs Rel <= 0.195, the efficiency gain is explained by pretraining or the data subset rather than by the proposed losses. If it stays near 0.208, the gain can be attributed to stereo supervision, but the '3/5 data' claim would still be invalid because the right-image supervision signal is not counted in the sample budget; the paper should then report the fraction 16,384/40,109 = 0.41 and restate the claim as 'using stereo pairs as extra supervision,' not 'using 3/5 of the data.'","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's advertised contribution—'competitive results with the baseline by only 3/5 times training data' (abstract, conclusion)—rests on Table 1, where Ours (Abs Rel 0.195) beats Zhou et al. (0.208). A data-efficiency conclusion requires that the comparison differ only in the training-set size. This condition fails. Section IV.A says the model was 'first pre-trained on the larger Cityscapes dataset' and then fine-tuned on KITTI, while the Zhou et al. baseline was trained on KITTI alone. At every training step the proposed method also uses the synchronized right stereo frame as an additional reconstruction target (Section III.A, Eq. 2 and Eq. 8), so each sample carries more supervision than the monocular sample used by the baseline; the limitations section concedes that 'our method requires rectified and temporally aligned stereo pairs during training.' The efficiency claim therefore cannot be isolated: the improvement could come from extra stereo supervision or pretraining rather than from needing fewer samples. The arithmetic is also internally inconsistent: Section IV.A reports 16,384 training sequences versus 40,109 for the baseline, which is 40.9% (about 2/5), while the abstract says 3/5 and the introduction says half. The paper does not state its own efficiency number consistently, and no ablation separates the proposed losses from the added supervision signals.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an unsupervised monocular video depth estimation method that extends the ego-motion-based view synthesis framework of Zhou et al. by adding a stereo left-right image reconstruction loss, a left-right disparity consistency loss, and an edge-aware smoothness term. The authors claim that the method achieves competitive results with the baseline using only 3/5 of the training data, and that it even outperforms some supervised methods. Experiments on KITTI and Make3D are reported, with KITTI results showing Abs Rel 0.195 versus 0.208 for the Zhou et al. baseline.","tokens_in":11517,"tokens_out":4003,"duration_ms":36692,"significance":"If the training-efficiency claim were convincingly supported, the paper would be a meaningful step toward reducing the data cost of unsupervised depth learning. The method combines two well-known ideas in a sensible way, and the reported KITTI results are directionally plausible. The strength of the paper is its use of a fully differentiable, standard geometric pipeline. However, the central efficiency claim is not supported by the current experimental design, and the arithmetic inconsistencies further weaken the contribution. The work could be substantial if the authors provide controlled comparisons that isolate the effect of the proposed losses from the extra stereo supervision and pretraining.","major_comments":[{"comment":"The central claim that the method achieves competitive results with the baseline using only 3/5 of the training data is not supported because the comparison is confounded. The proposed method uses the right stereo image as an additional reconstruction target (Eqs. 2 and 8) and is pre-trained on Cityscapes, whereas the Zhou et al. baseline is trained on monocular KITTI sequences only. The improvement seen in Table 1 could therefore be due to the extra stereo supervision or pretraining rather than to a genuine need for fewer samples. The limitations section explicitly concedes that the method requires rectified and temporally aligned stereo pairs for training. A controlled experiment, such as training the baseline with the same stereo losses or training the proposed method without stereo supervision and without pretraining, is needed to substantiate the efficiency claim; without it, the claim collapses.","section":"Abstract; Section IV.A; Table 1; Section IV.B (Limitations)"},{"comment":"The numerical claim about training data is internally inconsistent. Section IV.A reports 16,384 training sequences versus 40,109 for the baseline, which is approximately 41% (about 2/5), while the abstract and conclusion state 3/5, the introduction states 'half,' and Section III.A mentions '1/2x.' The authors should correct this arithmetic and state the efficiency claim consistently with the actual experimental protocol.","section":"Section IV.A; Abstract; Introduction; Section V (Conclusion)"},{"comment":"The definition of the disparity consistency loss is unclear. The symbols d_l^t and d_r^t are predicted disparities, but the notation d̂_l^t and d̂_r^t is not precisely defined; the text says they are 'computed by the corresponding disparity map' without specifying the warping operation. For reproducibility, the authors should clearly state that d̂_l^t is the left disparity obtained by sampling d_r^t using the right-to-left warping and vice versa, consistent with Godard et al. [10].","section":"Section III.A, Eq. (3)"}],"minor_comments":[{"comment":"The paper reports a training time of roughly 32 hours on a single 1080Ti GPU but does not provide the baseline's training time or the number of parameters, making the efficiency claim even harder to evaluate.","section":"Section IV.A"},{"comment":"The numerical results for the Zhou et al. baseline differ between Table 1 and Table 2 (e.g., Abs Rel 0.208 vs 0.207, RMSE 6.856 vs 6.658). The authors should ensure that the baseline numbers are consistent across tables and clearly state the source of each set of numbers.","section":"Table 1 and Table 2"},{"comment":"The claim that the method 'even outperforms the classical supervised methods' is overstated: on Sq Rel and RMSE, the supervised results of Eigen et al. and Liu et al. are better than the proposed method's results.","section":"Section IV.B"},{"comment":"The manuscript contains numerous typos and grammatical errors (e.g., 'touch to acquire,' 'dept,' 'un' in the abstract; 'This results is trained' in Section IV.A). A thorough copy-editing pass is needed.","section":"Abstract; throughout"},{"comment":"The sentence 'This results is trained in a total of 22,801 sequences where 16,384 for training and 5,730 for validation' is ungrammatical and unclear about the origin of the 22,801 total; please rephrase.","section":"Section IV.A"}],"recommendation":"major_revision","confidential_remarks":"The paper builds on a well-known baseline and a well-known stereo consistency loss; the claimed contribution of training efficiency is the main selling point. If the authors cannot provide controlled experiments that isolate the effect of the proposed losses from the extra stereo supervision and Cityscapes pretraining, the efficiency claim should be removed or substantially weakened. Even after revision, the novelty relative to Godard et al. is modest, and the editors may want to weigh whether the incremental contribution meets the journal's bar."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a modest engineering paper that adds a stereo reconstruction term and left-right consistency to Zhou et al.'s unsupervised depth-from-video framework, with a small gain on KITTI. The central novelty claim is not supportable, and the '3/5 training data' claim is a confounded and internally inconsistent comparison.\n\nThe geometry is standard, the implementation seems reasonable, and the reported numbers are plausible. The paper also includes a Make3D generalization test and a comparison to two supervised baselines, which is fine but not decisive.\n\nThe main issue is that the claimed contribution was already published: Monodepth2 (Godard et al., arXiv:1806.01260) combines temporal photometric loss with stereo left-right consistency, and Zhan et al. did the same in 2018. The paper's assertion that no previous approach exists is simply wrong. The training-efficiency claim is also misleading: the method uses stereo pairs as extra supervision and is pre-trained on Cityscapes, so the comparison with Zhou et al. does not isolate the effect of fewer samples. The arithmetic is inconsistent—16,384 vs 40,109 is about 41%, which matches '2/5' more than '3/5'—and the abstract says 3/5 while the introduction says half. The paper also lacks ablations, error bars, and a comparison to the very methods it is combining.\n\nThere is also an unfinished sentence ('We guess this is because of') and a stray section heading 'Resnet 18 and Resnet 50', which suggests the manuscript was not carefully polished.\n\nI would not send this to peer review at a serious venue. It is an incremental improvement that may work, but the claims are not positioned honestly, and the missing citations are disqualifying. A careful reader might use it as a data point, but it does not deserve referee time.","headline":"An incremental stereo-plus-temporal depth estimator whose headline claims are undercut by missing prior work and a confounded efficiency comparison.","tokens_in":12158,"tokens_out":4398,"would_cite":false,"duration_ms":39628,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Stereo-pair reconstruction lets unsupervised depth networks train on far fewer videos","keywords":["unsupervised depth estimation","monocular video","ego-motion","stereo disparity","left-right consistency","view synthesis","self-supervised learning","scene depth"],"falsifier":"Train the proposed method and the temporal-only baseline on the exact same number of video frames and stereo pairs, with the same pre-training and the same optimizer, then compare depth error on the standardized test split; if the proposed method needs as many frames as the baseline when the extra supervision is controlled for, the fewer-data claim is falsified.","tokens_in":11004,"feed_emoji":"📷","tokens_out":6032,"duration_ms":60617,"temperature":0.7,"pith_summary":"This paper tries to show that an unsupervised monocular depth estimator can learn from unlabeled video plus ordinary stereo pairs, and that the extra view-reconstruction signal makes training much more data-efficient. The method builds on an existing temporal view-synthesis approach and adds two stereo-based supervision terms: reconstructing the left and right images from the predicted disparities, and enforcing consistency between the left and right disparity maps. On an autonomous driving benchmark, the proposed network reaches a mean absolute relative error of 0.195, better than the 0.208 of the temporal-only baseline and better than some supervised methods, while being trained on roughly 16,384 sequences instead of the baseline's 40,109. If this holds, depth estimation for driving scenes needs no ground-truth labels and substantially less unlabeled footage.","feed_headline":"Stereo reconstruction lets depth networks learn from far fewer videos","feed_subtitle":"Adding right-eye view and disparity consistency to ego-motion training improves accuracy and cuts data needs.","key_machinery":"The load-bearing object is the combined view-synthesis loss that sums the temporal ego-motion reconstruction term, the stereo left-right image reconstruction term, and the disparity-consistency term. The stereo terms are what make the method data-efficient: they convert each training sample into two additional reconstruction targets without needing labels, so the depth network receives denser per-pixel supervision. Differentiable bilinear sampling lets gradients flow through both warps, and an explainability mask plus an edge-aware smoothness term disciplines the otherwise under-constrained problem.","core_discovery":"The central claim is that combining two complementary image-reconstruction tasks in one loss — temporal view synthesis driven by ego-motion and spatial stereo reconstruction driven by disparity — gives a single network that outputs a dense depth map from one image, with no ground-truth depth and no known pose at test time. During training the left frame is reconstructed both from neighboring video frames and from the right stereo frame; the network simultaneously predicts left and right disparities, and a left-right consistency term forces them to agree. An edge-aware L2 smoothness penalty encourages piecewise-smooth depth without blurring object contours. The reported outcome is that this combined supervision improves depth accuracy (Abs Rel 0.195 versus 0.208 for the temporal-only baseline), reduces the amount of training video needed to roughly 60 percent of the baseline, and transfers to a different outdoor dataset without retraining.","pith_inferences":["A conservative reading is that the reported speed-up is not yet isolated: the proposed model also consumes stereo pairs and is pre-trained on a larger city-driving dataset, so the fewer-data comparison mixes new losses with extra input signal and initialization.","The recipe suggests a general principle: any additional geometric view — another camera, a known pose, or a second time span — can be added as a reconstruction target to shrink the data needed by self-supervised depth learners.","A testable extension would apply the same disparity-consistency idea to unrectified or unknown-intrinsic video by predicting a lightweight epipolar correction, which the paper explicitly leaves to future work."],"forward_implications":["If the claim is correct, an autonomous-driving depth system can be trained from ordinary dashcam video plus rectified stereo pairs, with no labeled depth and no measured ego-motion.","The same extra-stereo supervision should produce usable depth on scenes the network has not seen, since the model trained on one dataset transfers to another without retraining.","The training-data savings are large enough that a single GPU can train the model in tens of hours rather than requiring massive unlabeled collections.","Because the method learns one network with a fixed scale from stereo geometry, it avoids the global scale ambiguity that plagues purely monocular ego-motion depth estimators."],"supporting_citations":[{"why":"Provides the temporal ego-motion depth and pose learning baseline that this method extends, and supplies the comparison for the fewer-data claim.","marker":"[14]"},{"why":"Introduces the left-right consistency and stereo reconstruction losses whose supervision this method imports into video depth learning.","marker":"[10]"},{"why":"Supplies the train/test split and supervised depth baselines against which the method reports competitive accuracy.","marker":"[4]"},{"why":"Supplies the differentiable bilinear sampling used to warp source views into the target frame.","marker":"[20]"},{"why":"Provides the encoder-decoder architecture with multi-scale disparity prediction that the depth network is based on.","marker":"[19]"},{"why":"Provides the larger city-driving dataset used to pre-train the network before fine-tuning, relevant to the data-efficiency comparison.","marker":"[25]"}],"fun_headline_variants":["Stereo consistency cuts depth-training video needs by 40%","Left-right disparity consistency improves depth accuracy from video","Ego-motion plus stereo disparity cuts training data by 40%","One network, two reconstructions: better depth with less data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central comparison assumes the extra stereo losses are what allow training on less data, but the proposed pipeline also uses stereo pairs as additional input and a larger pre-training dataset, so the fewer-frames claim has not been shown in isolation.","fun_headline_variants_meta":{"raw":{"variants":["Stereo consistency cuts depth-training video needs by 40%","Left-right disparity consistency improves depth accuracy from video","Ego-motion plus stereo disparity cuts training data by 40%","One network, two reconstructions: better depth with less data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001241,"raw_usage":{"total_tokens":5091,"prompt_tokens":942,"completion_tokens":4149,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":558,"completion_tokens_details":{"reasoning_tokens":4079}},"tokens_in":558,"tokens_out":4149,"duration_ms":28940,"temperature":1.0,"reasoning_tokens":4079,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:28:36.045090+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train the proposed method and the temporal-only baseline on the exact same number of video frames and stereo pairs, with the same pre-training and the same optimizer, then compare depth error on the standardized test split; if the proposed method needs as many frames as the baseline when the extra supervision is controlled for, the fewer-data claim is falsified.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the temporal ego-motion depth and pose learning baseline that this method extends, and supplies the comparison for the fewer-data claim."},{"cited_title":"Godard, O","cited_arxiv_id":null,"evidence_quote":"Introduces the left-right consistency and stereo reconstruction losses whose supervision this method imports into video depth learning."},{"cited_title":"Eigen, C","cited_arxiv_id":null,"evidence_quote":"Supplies the train/test split and supervised depth baselines against which the method reports competitive accuracy."},{"cited_title":"Mayer, E","cited_arxiv_id":null,"evidence_quote":"Provides the encoder-decoder architecture with multi-scale disparity prediction that the depth network is based on."},{"cited_title":"Cordts, M","cited_arxiv_id":null,"evidence_quote":"Provides the larger city-driving dataset used to pre-train the network before fine-tuning, relevant to the data-efficiency comparison."}],"review_version":1}