{"id":"8157fee8-dc8a-4fcf-85ee-cd3949d8c745","arxiv_id":"1908.02126","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A two-discriminator adversarial loss improves monocular depth estimation when labeled image-depth pairs are scarce, using unlabeled images during training.","lead":"This paper proposes a semi-supervised adversarial training method for monocular depth estimation, using a generator plus two discriminators, and tests it on standard benchmarks with limited labeled data. It reports consistent error reductions and improved domain adaptation, though several quantitative details need independent verification.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Small semi-supervised gains with no repeated-seed statistics: Make3D rel 0.158 vs 0.153 and cross-table inconsistencies make the unlabeled-data benefit indistinguishable from run-to-run noise.","rationale":"The reader's weakest assumption about depth-discriminator degeneracy is plausible but not the first threat: the DD-only ablation already gives a gain over L1 (rel 0.191 vs 0.201), so a degenerate DD is not obviously the failure mode. The immediate threat is that the headline gains are within run-to-run noise. The paper reports a single run per configuration, no confidence intervals, and several cross-table inconsistencies. Because the claimed semi-supervised improvement is small (Make3D 0.005 rel; NYU Table 2 best supervised 0.192 to semi 0.183 is larger but still one number), the causal role of unlabeled data is not statistically secured. This supports the reader's CONDITIONAL verdict: require seed-variance evidence before the claim is accepted. I do not see a fatal flaw; the method is clearly specified at the conceptual level, and the limitations section is honest about the saturated-data regime.","tokens_in":25107,"tokens_out":8426,"duration_ms":88116,"concrete_test":"Rerun the Make3D Ours-GAN-supervised vs Ours-GAN-semi comparison and the NYU 500-label L1 vs full semi-supervised comparison with at least 5 independent seeds and identical hyperparameters; report mean and std for rel, RMSE, and log10. If the mean semi-supervised gain is less than one standard deviation or the ordering flips, the central claim is not established. In the same runs, reproduce Table 2's edge-aware row and Table 4's Hu et al. 500-label row under identical protocol to determine whether they are the same setting; any protocol difference must be stated in the paper.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that unlabeled RGB images, through the two-discriminator loss, drive the limited-label gains. The cleanest evidence is the Make3D comparison: Ours-GAN supervised rel 0.158/rms 6.139 vs Ours-GAN semi rel 0.153/rms 6.054, and the KITTI comparison 0.107/4.405 vs 0.093/4.195. These differences are small, and no error bars or repeated-seed statistics are reported anywhere. The paper itself contains same-setting numbers that differ by comparable or larger amounts: Table 2 lists the 'Edge Aware [3] supervised' 500-label baseline as rel 0.201 / rms 0.750 / log10 0.083, while Table 4 lists 'Hu et al. [3] 500' as rel 0.197 / rms 0.837 / log10 0.084. With no variance estimate, a 0.005 rel gain on Make3D is not distinguishable from run-to-run noise or minor protocol variation. This is load-bearing: if the gain is noise, the paper supports 'GAN losses help supervised depth training,' but not the stated semi-supervised claim that unlabeled images improve accuracy.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a semi-supervised adversarial framework for monocular depth estimation that uses a small number of labeled image-depth pairs together with many unlabeled RGB images. A generator (backbone can be U-Net, FCRN, or the Hu et al. network) regresses a depth map; two discriminators then evaluate the output: a pair discriminator (PatchGAN) on the image-depth pair and a depth discriminator on the depth channel alone. The generator loss combines the feedback from both discriminators, plus an L1 regression term on labeled pairs. Experiments on NYU Depth, Make3D, and KITTI, including a KITTI-to-Make3D domain adaptation setting, report improvements over supervised baselines and state-of-the-art results on Make3D with limited labels.","tokens_in":25252,"tokens_out":9684,"duration_ms":90407,"significance":"If the reported results hold, the contribution is practically useful: a relatively simple way to exploit unlabeled RGB images in depth regression when labeled depth is scarce, with an architecture that can wrap several existing depth networks. The paper includes useful ablations over loss functions, backbones, labeled-data counts, and the number of additional unlabeled images, and it explicitly discusses the regime where semi-supervised learning stops helping. However, the absence of error bars or repeated-seed statistics, the small size of several key gains, and some internal inconsistencies in the core equations prevent the central claim from being fully established as written.","major_comments":[{"comment":"The paper reports no error bars, no results over multiple random seeds, and no significance tests anywhere in the experimental section. The key semi-supervised improvements are small in absolute terms (Make3D rel 0.158 to 0.153 in Table 6; KITTI rel 0.107 to 0.093 in Table 7), and GAN training is known to be sensitive to initialization. Moreover, Table 4 shows the opposite direction for one configuration (Unet at 1000 labels: supervised rel 0.365 vs semi-supervised 0.369). Without a variance estimate, the reader cannot tell whether the reported gains reflect the method or run-to-run noise. This is load-bearing for the central claim that unlabeled images improve depth accuracy; please provide mean and standard deviation over at least three runs for the main comparisons, and ideally paired differences.","section":"Sec. 4.3-4.5, Tables 2, 4, 6, 7"},{"comment":"The formal definition of the proposed loss is internally inconsistent. In Eq. (5), the terms intended for the depth discriminator are written with PD: the third term is E[log PD(d)] and the fourth is E[log(1 - PD(G(i')))], whereas Eq. (3) defines the depth discriminator objective with DD. In Sec. 3.3, Eq. (9) labels the DD loss but uses PD in both expectations, and Eq. (8) mixes the labeled index i with the unlabeled index i'. Since the method is defined by these equations, the manuscript does not currently specify the algorithm unambiguously. Please correct all three equations and make the index conventions consistent throughout.","section":"Sec. 3.2 Eqs. (5); Sec. 3.3 Eqs. (8)-(9)"},{"comment":"The hyperparameter selection protocol is not described for the NYU and Make3D experiments. The text in Sec. 3.2 reports that lambda = 0.7 was found to give 'a reasonable result', and Sec. 3.3 mentions a larger beta at initial training, but no validation split or tuning procedure is given for these datasets. Only the KITTI section (Sec. 4.5) mentions a 160-sample validation set. If lambda, beta, or early stopping were chosen by inspecting test-set numbers, the reported results are optimistically biased. Please specify the validation protocol for each dataset, or fix the hyperparameters and the beta schedule a priori.","section":"Sec. 3.2-3.3, Sec. 4.3-4.4"},{"comment":"The depth discriminator is trained on only 400-500 real depth maps (NYU, Make3D), but no experiment checks whether it learns a useful, non-degenerate distribution. The DD-only ablation in Table 2 reaches rel 0.191, close to the L2 baseline, and there is no analysis of the discriminator's separability on held-out depth statistics or of its capacity. Because the depth discriminator supplies the prior p(d) in the claimed Bayesian interpretation, the authors should add an experiment varying the discriminator capacity or measuring its generalization, to show that the semi-supervised gain is not an artifact of memorizing the few labeled depth maps.","section":"Sec. 3.1.2, Eq. (3), Table 2"}],"minor_comments":[{"comment":"The phrase 'Different traditional GANs' loss' should read 'Different from traditional GANs' loss' for clarity.","section":"Sec. 1, Sec. 3.2"},{"comment":"Table 2 and Table 4 report different supervised numbers for the same 500-label backbone (Edge Aware [3] rel 0.201 vs Hu et al. [3] rel 0.197); please clarify which loss (edge-aware vs L1) was used for each row, since the current text in Sec. 4.3 could be read as inconsistent.","section":"Sec. 4.3, Tables 2 and 4"},{"comment":"The Cityscape images are described as 'drawn' from the dataset; please state explicitly that they are used without depth labels, and describe how they are sampled.","section":"Sec. 4.5"},{"comment":"The paper would benefit from releasing code and trained models; at present no reproducibility artifacts are provided, which makes it harder to verify the small semi-supervised gains.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of TPAMI and the proposed two-discriminator semi-supervised framework is a reasonable extension of adversarial training. However, the statistical evidence for the central claim is currently weak: most headline comparisons are single runs with small deltas, and one configuration in Table 4 goes in the opposite direction. The equation typos in Sec. 3.2-3.3 are easily fixable, but they should be corrected before further review. I would like to see the revised version with repeated-seed results and a clear hyperparameter validation protocol before recommending acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real empirical contribution: a two-discriminator adversarial loss, one on image-depth pairs and one on depth maps alone, helps monocular depth regression under small labeled sets and transfers across datasets. Second, the size of the claimed semi-supervised gain is not established by the numbers as reported, because there are no error bars and the paper's own baseline numbers shift between tables.\n\nWhat's new: D2GAN had two discriminators but unconditioned; here the pair discriminator conditions on RGB and the depth discriminator enforces the marginal depth prior. The Bayesian framing (p(d), p(I|d)) is informal but useful. The paper does good things: it ablates loss functions, backbones (Unet, FCRN, Hu), label counts (100/500/1000), and the number of unlabeled images; it tests domain adaptation from KITTI to Make3D; and it includes a limitations section that openly says the method stops helping when labels are abundant. That honesty is worth noting.\n\nSoft spots, in order. No variance estimates anywhere. On Make3D the semi-supervised gain is rel 0.158 to 0.153; on KITTI it is 0.107 to 0.093. Without repeated seeds, those differences are within the range of run-to-run noise for this kind of training. The stress-test note is right that Table 2 and Table 4 disagree on the same supervised 500-image baseline (rel 0.201 vs 0.197, rms 0.750 vs 0.837), which makes the gain hard to size. The KITTI row is also odd: rel 0.107 beats Kuznietsov's 0.108, but rms 4.405 is far worse than their 3.518; an unexplained protocol difference is plausible, but the paper needs to say what it is. There are obvious typos in Eqs. (5) and (9), where the depth discriminator is written with PD terms. Minor, but they should be fixed. The comparison set lacks the closest semi-supervised and unsupervised GAN baselines in the same small-label regime; Kuznietsov is discussed but never directly compared. And the depth discriminator trained on 400-500 real depth maps could be memorizing; no experiment varies its capacity or measures the quality of its signal. That concern is real, though the ablations give indirect evidence that the two discriminators together beat either alone.\n\nBottom line: the core idea is plausible and the paper is honest, but the evidence for the central semi-supervised claim is currently too weak to take at face value. A serious referee should ask for repeated-seed statistics, a consistent evaluation protocol, fixed equations, and direct baselines. I would send it to review, not desk reject it, but it needs major revision before the gains can be believed.","headline":"Useful two-discriminator recipe for semi-supervised depth, but the central gain is small and not yet backed by variance estimates or consistent baselines.","tokens_in":25932,"tokens_out":3127,"would_cite":true,"duration_ms":32280,"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":"The paper claims that monocular depth estimation can be made semi-supervised: with a few hundred labeled pairs, adversarial feedback from two discriminators lets unlabeled RGB images drive depth accuracy to state-of-the-art levels.","keywords":["monocular depth estimation","semi-supervised learning","generative adversarial networks","depth regression","unlabeled RGB images","domain adaptation","limited labeled data"],"falsifier":"On NYU v2 with 500 labeled pairs, train the full model but freeze the depth discriminator at initialization so it supplies no learned feedback; if accuracy stays at the combined-model level (rel 0.183) instead of falling toward the pair-discriminator-only level (rel 0.198), the depth discriminator is not carrying the claimed benefit.","tokens_in":24764,"feed_emoji":"📏","tokens_out":10585,"duration_ms":101851,"temperature":0.7,"pith_summary":"This paper tackles the cost of collecting aligned image-depth pairs for monocular depth estimation. It proposes a semi-supervised adversarial setup in which one generator regresses depth and two discriminators grade the result: one checks the image-depth pair, the other checks the depth map alone. The discriminators' feedback enters the generator's loss, so a large pool of easily available unlabeled RGB images can shape the predicted depth distribution without ground-truth depth. The paper reports that this approach lowers relative error by 28–43% on NYU v2 under limited labels, reaches state-of-the-art accuracy on the small Make3D dataset, and improves KITTI-to-Make3D domain adaptation.","feed_headline":"Two adversarial critics cut depth error with few labeled pairs","feed_subtitle":"Adding unlabeled RGB images to a few hundred labeled pairs beats fully supervised L1/L2 training on depth accuracy.","key_machinery":"The load-bearing object is a three-player adversarial game among one generator and two discriminators. The pair discriminator is a PatchGAN, a convolutional classifier that scores overlapping $70\\times70$ patches of the concatenated image and predicted depth; the depth discriminator is structurally similar but receives only depth maps. Their feedback reaches the generator through the weighted loss above, and the paper interprets the arrangement as a Bayesian decomposition $p(d|I)\\propto p(I|d)p(d)$, with the depth discriminator providing the prior $p(d)$ and the pair discriminator providing the joint $p(d,I)$.","core_discovery":"The central claim of the paper is that realism is a usable training signal for monocular depth regression when labeled image-depth pairs are scarce. The generator loss is $L_G = \\lambda \\mathbb{E}_{i'\\sim p_{data}(i')}[\\log(1-P_D(i',G(i')))] + (1-\\lambda) \\mathbb{E}_{i'\\sim p_{data}(i')}[\\log(1-D_D(G(i')))]$, with $\\lambda = 0.7$, and supervised iterations add an $\\ell^1$ regression term. The pair discriminator $P_D$ judges whether an image and its predicted depth form a real image-depth pair, while the depth discriminator $D_D$ judges whether a predicted depth map alone resembles real depth. With only hundreds of labeled pairs, the two discriminators let the generator exploit unlabeled RGB images; the paper reports rel 0.183 on NYU v2 with 500 labeled pairs versus 0.195 for supervised L2, and rel 0.153/0.158 on Make3D.","pith_inferences":["The depth discriminator is trained only on the labeled depth maps, so its contribution plausibly depends on how representative those maps are; ablating its capacity at fixed labeled count would test this, and the paper does not report such a study.","The same pair-plus-depth discriminator split could apply to other dense regression tasks with scarce labels, such as surface-normal estimation or image restoration, wherever a 'real output' distribution is well defined.","Since the unlabeled images contribute only realism feedback, combining this loss with geometric self-supervision, such as left-right or photometric consistency, might extract additional signal from the same unlabeled pool.","The saturation pattern suggests a practical rule: collect about ten unlabeled images per labeled pair before spending more on annotation, a rule other datasets could test directly."],"forward_implications":["With limited labeled pairs, adding unlabeled RGB images improves depth accuracy, and the benefit grows up to roughly ten unlabeled images per labeled pair before saturating.","When labeled pairs are abundant (beyond about 3.5K on NYU v2), the semi-supervised gain disappears, so the method targets the scarce-label regime rather than replacing large-scale supervision.","The same loss improves three different generator architectures, so the dual-discriminator feedback transfers across backbone designs.","The framework supports domain adaptation: a model trained on KITTI pairs plus Make3D RGB images cuts the KITTI-trained Make3D relative error from 0.555 to 0.447.","Adversarial feedback yields qualitatively sharper object boundaries in predicted depth maps than L1, L2, or Huber losses under limited labels."],"supporting_citations":[{"why":"Supplies the PatchGAN discriminator design and the conditional adversarial loss that the pair discriminator and generator loss are built on.","marker":"[39]"},{"why":"Provides the fully convolutional residual generator backbone and the berHu supervised baseline that semi-supervised training improves.","marker":"[2]"},{"why":"Provides the edge-aware baseline generator and supervised loss that the paper upgrades with the two-discriminator framework.","marker":"[3]"},{"why":"Establishes the multi-scale CNN depth estimation baseline and the standard evaluation metrics and KITTI split used in the experiments.","marker":"[1]"},{"why":"Supplies the Make3D dataset and the original superpixel-MRF baseline that the small-data comparisons build upon.","marker":"[9]"},{"why":"Supplies the NYU Depth v2 dataset, its official train/test scenes, and the semantic-region labels used in the error analysis.","marker":"[44]"},{"why":"Is the earlier GAN-based depth estimator whose supervised 12K-label result serves as the direct comparison for the supervised GAN loss.","marker":"[27]"},{"why":"Is the prior semi-supervised depth method whose requirement for registered stereo pairs contrasts with the present use of unlabeled RGB images alone.","marker":"[33]"},{"why":"Provides the deep ordinal regression baseline used in Make3D and KITTI-to-Make3D adaptation comparisons.","marker":"[11]"}],"fun_headline_variants":["Adversarial critics sharpen depth with sparse pairs","Dual discriminators teach depth from few images","Semi-supervised depth: realism beats more labels","Fewer pairs, higher depth accuracy via adversarial feedback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The depth discriminator can learn a useful distribution of real depth maps from only the labeled depth maps available, and its feedback improves rather than distorts the regressed depth scale and structure.","fun_headline_variants_meta":{"raw":{"variants":["Adversarial critics sharpen depth with sparse pairs","Dual discriminators teach depth from few images","Semi-supervised depth: realism beats more labels","Fewer pairs, higher depth accuracy via adversarial feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000624,"raw_usage":{"total_tokens":2916,"prompt_tokens":1002,"completion_tokens":1914,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":618,"completion_tokens_details":{"reasoning_tokens":1853}},"tokens_in":618,"tokens_out":1914,"duration_ms":14587,"temperature":1.0,"reasoning_tokens":1853,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:51:42.870355+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On NYU v2 with 500 labeled pairs, train the full model but freeze the depth discriminator at initialization so it supplies no learned feedback; if accuracy stays at the combined-model level (rel 0.183) instead of falling toward the pair-discriminator-only level (rel 0.198), the depth discriminator is not carrying the claimed benefit.","supporting_citations":[{"cited_title":"Image-to-image translation with conditional adversarial networks,","cited_arxiv_id":null,"evidence_quote":"Supplies the PatchGAN discriminator design and the conditional adversarial loss that the pair discriminator and generator loss are built on."},{"cited_title":"Deeper depth prediction with fully convolutional residual net- works,","cited_arxiv_id":null,"evidence_quote":"Provides the fully convolutional residual generator backbone and the berHu supervised baseline that semi-supervised training improves."},{"cited_title":"Revisiting single image depth estimation: toward higher resolution maps with accurate object boundaries,","cited_arxiv_id":null,"evidence_quote":"Provides the edge-aware baseline generator and supervised loss that the paper upgrades with the two-discriminator framework."},{"cited_title":"Depth map prediction from a single image using a multi-scale deep network,","cited_arxiv_id":null,"evidence_quote":"Establishes the multi-scale CNN depth estimation baseline and the standard evaluation metrics and KITTI split used in the experiments."},{"cited_title":"Make3d: Learning 3d scene structure from a single still image,","cited_arxiv_id":null,"evidence_quote":"Supplies the Make3D dataset and the original superpixel-MRF baseline that the small-data comparisons build upon."},{"cited_title":"Indoor seg- mentation and support inference from rgbd images,","cited_arxiv_id":null,"evidence_quote":"Supplies the NYU Depth v2 dataset, its official train/test scenes, and the semantic-region labels used in the error analysis."},{"cited_title":"Depth map prediction from a single image with generative adversarial nets,","cited_arxiv_id":null,"evidence_quote":"Is the earlier GAN-based depth estimator whose supervised 12K-label result serves as the direct comparison for the supervised GAN loss."},{"cited_title":"Semi-supervised deep learning for monocular depth map prediction,","cited_arxiv_id":null,"evidence_quote":"Is the prior semi-supervised depth method whose requirement for registered stereo pairs contrasts with the present use of unlabeled RGB images alone."},{"cited_title":"Deep ordinal regression network for monocular depth estimation,","cited_arxiv_id":null,"evidence_quote":"Provides the deep ordinal regression baseline used in Make3D and KITTI-to-Make3D adaptation comparisons."}],"review_version":1}