{"id":"2db349fe-c209-4894-bec6-c9808e216470","arxiv_id":"2505.08178","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A depth-guided, occlusion-aware refinement network with a temporal optical-flow loss slightly reduces disparity error on laparoscopic images.","lead":"This paper presents a neural network that cleans up depth maps from stereo surgery cameras by borrowing depth information from a single-camera estimate and by using video motion as extra training signal. It reports slightly lower depth errors than prior models on the SCARED laparoscopic dataset.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (2) contradicts the stated occlusion-mask semantics: with M approaching 1 denoting occlusion, the output in occluded regions is the coarse disparity S, not the refined inverse depth D_inv, so the central depth-guidance mechanism is absent from the model as written.","rationale":"The reader's named weakest assumption was LeReS domain bias, which is plausible and is partly conceded in Section 4.1. The Eq. (2) mask-semantics mismatch is a more direct threat to the central claim because it concerns whether the published architecture can implement depth-guided occlusion refinement at all. I mark agreement as partial: the reader listed this equation in the rationale but did not make it the weakest assumption. The concern reinforces the CONDITIONAL verdict rather than overturning it, because a corrected equation or a released code artifact could resolve it; without that, the Table 1 gains cannot be confidently attributed to the proposed mechanism. The marginal improvements, missing error bars, and absent code noted by the reader remain additional reasons not to accept as-is, so I keep the reader's verdict unchanged.","tokens_in":6885,"tokens_out":10311,"duration_ms":97526,"concrete_test":"Retrain the DGOR-LG prototype on the same SCARED split with Eq. (2) replaced by S_hat = (1-M)*S + M*D_inv_hat, the form that makes M approaching 1 select depth in occluded regions, holding all other hyperparameters fixed, and compare SEF and SEK EPE, Bad3, and RMSE with Table 1. If the numbers match Table 1 within a small tolerance, the reported model effectively used the corrected semantics and the text is only misworded; if they differ materially, the headline results depend on an undocumented weighting. As a diagnostic, also record the predicted M at LRC-occluded borders to confirm which term is active there.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that DGORNet refines occluded disparity using monocular depth because depth is unaffected by occlusion. Section 2.1 defines M in [0,1] with M approaching 1 as occluded and M approaching 0 as occlusion-free, but Eq. (2) is S_hat = M*S + (1-M)*D_inv_hat. Under that definition, occluded pixels receive the coarse disparity S and only occlusion-free pixels receive the depth-refined value. That is the opposite of the advertised depth-in-occlusion mechanism. Since L_s(S_hat, S_gt) and L_DC(M, M_LRC) supervise these quantities, the network cannot learn the intended behavior unless the implementation silently uses (1-M)*S + M*D_inv_hat or reverses the mask convention. No code is released, so the reader cannot tell which version produced Table 1. The paper's own Section 4.1 admission that LeReS outliers raise Bad3 is a separate reliability concern; Eq. (2) is more basic because even an ideal depth predictor would not be selected in occluded regions under the published formula. This is fixable as a typo or mask-convention error, but it must be resolved before the headline result can be attributed to depth-guided occlusion refinement.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DGORNet, a post-processing module that refines coarse disparity maps from an existing stereo matcher by combining them with monocular inverse-depth estimates from LeReS. It introduces an occlusion-mask prediction module trained with left-right-consistency-derived masks, a position embedding to provide spatial context, and an optical flow difference loss for semi-supervised temporal regularization on unlabeled frames. Experiments on the SCARED dataset (SEF and SEK) compare DGORNet attached to MSDESIS and GMStereo against four baselines, reporting best EPE 4.02 px and RMSE 4.93 mm, with an ablation study for the position embedding and OFDLoss components.","tokens_in":7168,"tokens_out":7370,"duration_ms":71504,"significance":"The paper addresses a relevant problem in laparoscopic disparity estimation, and the experimental design of attaching the same refinement module to two different base predictors is a clean way to measure transferability. The reported low overhead of the refinement module (3.83 s per 1000 predictions) is also a useful engineering property. If the claimed results are reproducible, DGORNet would be a sensible low-cost component for stereo laparoscopic disparity estimation. However, the advertised mechanism is currently undermined by the sign convention in Eq. (2), and the quantitative support is weakened by small, mixed margins and the absence of variance estimates.","major_comments":[{"comment":"The mask convention contradicts the central mechanism. The text states M in [0,1], with M→1 indicating occlusion and M→0 occlusion-free, but Eq. (2) computes S_hat = M*S + (1-M)*D_inv_hat. Under the stated semantics, occluded pixels keep the coarse disparity S, and only occlusion-free pixels receive the depth-refined inverse depth. Since Eq. (6) supervises M with the LRC-derived mask M_LRC, the learned mask will be near 1 exactly where the paper claims depth guidance should act, so the final prediction excludes D_inv_hat from occluded regions. If the intended formula is S_hat = (1-M)*S + M*D_inv_hat, or if the mask convention is reversed, this must be corrected and verified in the implementation; without released code, Table 1 cannot be attributed to the described depth-guided occlusion refinement.","section":"Section 2.1, Eq. (2)"},{"comment":"The reported advantage over the base predictors is small and inconsistent. On SEF, DGOR-LG changes EPE from 4.04 to 4.02 pixels but increases Bad3 from 43.95% to 45.37%, and DGOR-LM increases Bad3 from 46.04% to 47.04% while improving EPE by 0.34 pixels. The RMSE gains are 0.05–0.10 mm. No error bars, repeated-seed results, or statistical tests are reported for any table, and the ablation in Table 2 shows differences as small as 0.003 pixels EPE (2.207 vs 2.204) and a Bad3 change in the wrong direction for OFDLoss (13.29 vs 13.30). These margins are within typical run-to-run variation, so the claim that DGORNet outperforms state-of-the-art methods is not yet established.","section":"Section 4.1, Table 1"},{"comment":"The assumption that LeReS provides depth unaffected by occlusion is not validated for laparoscopic scenes. LeReS is pretrained on natural images, and the paper itself attributes the Bad3 increase to potential outliers introduced by the depth predictor LeReS. Because Eq. (1) applies only a learned per-pixel affine correction to the inverse depth, it is unlikely to remove spatially varying bias caused by tissue texture, specular highlights, or instruments. A quantitative evaluation of LeReS (e.g., error maps against SCARED ground truth, stratified by occlusion status) or an ablation with a domain-adapted depth predictor is needed before the central claim can be accepted.","section":"Section 3.2 and Section 4.1"},{"comment":"The OFDLoss as written is not a valid non-negative loss: L_flow = w_flow · (Δ_x F^k - Δ^k_x P) with positive weights w_flow can be negative, and minimizing it may encourage the signed difference to become more negative rather than closer to zero. If the intended quantity is an L1 or squared error, an absolute value or square must appear in Eq. (4). This matters because OFDLoss is one of the two claimed contributions and is used directly in Eq. (6).","section":"Section 2.3, Eq. (4)"}],"minor_comments":[{"comment":"The columns for Params and Runtime are not aligned; entries such as '0.33 (55.49)' and '3.83 (40.42)' need explicit labels (e.g., refinement-only versus total runtime) so the reader can interpret the reported numbers.","section":"Table 1"},{"comment":"The caption states 'five state-of-the-art methods' while the figure shows four panels (DeepPruner, HSM, GMStereo, MSDESIS); the caption and the text should be reconciled.","section":"Fig. 2 caption"},{"comment":"The notation for frame indices is inconsistent: the text refers to P^{t-1}_R while Eq. (3) and the surrounding discussion use P^{k-1}; please use a single consistent index.","section":"Section 2.3"},{"comment":"The notation 'clamp0,1' is undefined; write clamp(x, 0, 1) or min(1, max(0, x)) to specify the clipping operation.","section":"Eq. (5)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of the journal, but the novelty is incremental and the empirical evidence is thin. The Eq. (2) issue appears fixable as a typo or mask-convention error, but it directly affects the interpretation of the headline result. I would be willing to review a revision that resolves the mask contradiction, clarifies the OFDLoss definition, and adds variance estimates or repeated-seed statistics."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable plug-in idea for refining stereo laparoscopic disparity with monocular depth, but the paper as written has a load-bearing internal contradiction. Equation (2) blends coarse disparity S and refined inverse depth D_inv_hat with the occlusion mask M, where M→1 means occluded. That gives occluded pixels the coarse S and occlusion-free pixels the depth-refined value. This is the opposite of the advertised mechanism from the abstract, which says depth 'unaffected by occlusion' refines occluded regions. Unless the implementation silently flips the mask convention, the network cannot be learning what the paper claims. No code is released, so the reader can't tell which variant produced Table 1.\n\nWhat is genuinely useful: the overall setup—depth predictor plus disparity predictor plus refinement with position embeddings and a semi-supervised optical-flow difference loss—is a sensible combination for label-scarce surgical video. The base-predictor comparison (MSDESIS, GMStereo, HSM, DeepPruner) is the right experimental design, and the ablation on SEK isolates the PE and OFDLoss contributions. The computational overhead of the refinement is small (3.83 s/1000 predictions), which is a practical point.\n\nWhere it gets soft: the effect sizes are tiny—EPE on SEF goes from 4.04 to 4.02 for GMStereo, and Bad3 actually gets slightly worse. There are no error bars or repeated runs, so we can't tell if the gains are real. The LeReS depth pretrained on natural images may not transfer well to laparoscopy; the paper concedes this can raise Bad3. And the Eq. (2) issue isn't minor rhetoric—it's the center of the method.\n\nWho would get value from this: people working on stereoscopic depth refinement for endoscopy, and anyone testing 'depth helps disparity in occlusion' hypotheses. But the numbers should not be taken at face value until the equation is fixed or clarified.\n\nRecommendation: send it to peer review, but only with a firm instruction that the authors correct or explain Eq. (2), release code, and ideally report variance over multiple training runs. The idea is worth a referee's time; the current write-up is not yet an accurate report of its own method.","headline":"A plausible refinement plug-in whose central fusion equation appears to invert its own occlusion semantics; the reported gains are tiny and the code is absent.","tokens_in":7682,"tokens_out":3173,"would_cite":false,"duration_ms":29017,"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":"Monocular depth, which occlusions cannot touch, can repair disparity errors in stereo laparoscopic images.","keywords":["disparity estimation","disparity refinement","semi-supervised learning","occlusion","monocular depth guidance","laparoscopic images","optical flow difference loss","position embedding"],"falsifier":"A reader could settle the claim by evaluating DGORNet on a laparoscopic sequence with dense ground truth and computing EPE separately inside and outside occlusion masks: if the depth-guided refinement does not reduce EPE inside occlusion masks compared with the coarse disparity, the central premise is not carrying the gain. A stricter test is to corrupt the monocular depth with a known non-affine bias—such as a depth-dependent drift—and observe whether the final EPE degrades, since Eq. (1) can absorb only scale and shift errors.","tokens_in":1684,"feed_emoji":"🩺","tokens_out":6795,"duration_ms":101758,"temperature":0.7,"pith_summary":"This paper tries to establish that the hard failure regions of stereo matching in laparoscopic surgery—occlusions and texture-less tissue—can be repaired by borrowing depth from a monocular estimator, because a single camera never suffers the left-right occlusion that breaks stereo correspondence. To make that work, the authors propose DGORNet, a refinement network that predicts an occlusion mask, applies a learned scale-and-shift to the monocular inverse depth, and blends the corrected depth into the coarse disparity exactly where the mask says occlusion occurs. They add a Position Embedding module so the network knows where pixels sit in the image, and a semi-supervised Optical Flow Difference Loss that uses consecutive video frames to constrain disparity without dense labels. On the SCARED benchmark the refined outputs reach the best reported EPE (4.02 px) and RMSE (4.93 mm) on the full-video evaluation set, and the refinement itself costs only about 3.83 seconds per 1000 predictions. If correct, this gives laparoscopic stereo pipelines a small post-processor that can upgrade an existing matcher without retraining it.","feed_headline":"Depth-guided refinement hits 4.02-pixel error in laparoscopic stereo","feed_subtitle":"A small mask-weighted refiner fixes occluded tissue regions with monocular depth, no dense labels needed.","key_machinery":"The load-bearing mechanism is the mask-weighted blend of a scale-and-shift corrected monocular inverse depth with the coarse stereo disparity. The Occlusion Mask Prediction Module (OPM) takes the left image, a position map, coarse depth, and coarse disparity, and emits a soft occlusion mask. The Disparity Refinement Module (DRM), a 2D hourglass network, emits scale map $K$ and shift map $B$, so Eq. (1) converts the monocular inverse depth into a refined inverse depth, and Eq. (2) blends it with the coarse disparity according to the mask. The Position Embedding adds sinusoidal and cosine coordinate encodings to the OPM input so that error-prone border and edge regions are localized. The Optical Flow Difference Loss uses the x-component difference between left and right optical flow fields across consecutive frames, with the y-component difference as an adaptive confidence weight, to constrain the refined disparity on unlabeled frames.","core_discovery":"DGORNet's central claim is that monocular depth is an occlusion-free signal that can fix the disparity map exactly where stereo matching fails. The network takes a coarse disparity from an existing stereo matcher, a coarse inverse depth from a monocular depth predictor, the left image, and a position embedding; it then predicts a soft occlusion mask $M$, and maps the inverse depth to a refined inverse depth $\\hat{D}_{inv} = K D_{inv} + B$ via learned scale and shift maps. The final disparity is $\\hat{S} = M S + (1-M) \\hat{D}_{inv}$, so the mask decides where to trust the original stereo result and where to trust the depth-guided correction. On SCARED, the authors report that this beats both the base stereo predictors and four state-of-the-art methods on EPE and RMSE, with the best numbers 4.02 px and 4.93 mm on the full-video evaluation set. The slightly higher Bad3 on that set is attributed by the authors to outliers introduced by the monocular depth predictor, a limitation they state directly.","pith_inferences":["Extension: the paper's explanation predicts that the EPE gain should concentrate inside occlusion masks; measuring EPE separately inside and outside predicted occlusion regions on a different laparoscopic dataset would test whether depth guidance, not some other component, is carrying the improvement.","Extension: because Eq. (1) only corrects for scale and shift, a monocular depth estimator with a non-affine systematic bias—for example one that misreads specular highlights or surgical instruments—would inject structured errors that the learned correction cannot absorb.","A testable consequence the authors leave implicit: replacing the left-right-consistency-derived occlusion masks with ground-truth occlusion masks during training should strengthen the mask branch and therefore the final disparity, since the current supervision inherits the limitations of LRC checks.","The runtime numbers suggest a practical deployment path: the small prototype runs in about 40 seconds per 1000 predictions and is described as suitable for latency-sensitive settings, so the refinement is plausible as a real-time intraoperative component."],"forward_implications":["Existing stereo matchers for laparoscopic images can be upgraded by appending DGORNet as a small post-processor, without retraining the base matcher.","Semi-supervised training with video frames can reduce the need for dense disparity labels, since unlabeled frames are regularized by optical-flow temporal consistency.","Depth-guided refinement should be most beneficial exactly at occluded object boundaries, tool edges, and image borders, where stereo matching errors concentrate.","The refinement overhead is small relative to the base matcher's runtime, so the accuracy gain can be obtained under latency constraints by choosing the smaller base predictor.","If the depth signal really is occlusion-free, the same refinement idea could be transferred to other stereo settings where occlusion is the dominant error source."],"supporting_citations":[{"why":"Supplies the SCARED dataset, its ground-truth disparity, and the RGB-video frames used for both training and the SEF/SEK evaluation splits.","marker":"[1]"},{"why":"Provides the occlusion-type classification and left-right consistency check that the paper uses to derive ground-truth occlusion masks for supervision.","marker":"[4]"},{"why":"Serves as the MSDESIS disparity predictor used in the small prototype and as the source of the ConvBnLRelu building block used in the OPM.","marker":"[8]"},{"why":"Is cited as the established practice that motivates using a monocular depth predictor (LeReS) to supply the inverse depth guidance.","marker":"[10]"},{"why":"Provides the GMStereo disparity predictor used in the large prototype and the GMFlow optical-flow estimator used for the OFDLoss.","marker":"[15]"},{"why":"Supplies the sinusoidal and cosine position-encoding idea that the Position Embedding module adapts for spatial context.","marker":"[5]"},{"why":"Supplies the 2D hourglass architecture used in the Disparity Refinement Module to predict the scale and shift maps.","marker":"[2]"}],"fun_headline_variants":["Monocular depth fixes occluded stereo in laparoscopy","Depth-guided refiner cuts laparoscopic disparity error","Semi-supervised refiner uses depth to see past occlusions","Occlusion-aware disparity via monocular depth, no dense labels","DGORNet: depth guidance beats stereo baselines in surgery"],"cache_read_input_tokens":9856,"weakest_assumption_plain":"The load-bearing premise is that the off-the-shelf monocular depth estimator remains accurate in laparoscopic scenes after a learned scale-and-shift correction; if its depth is systematically wrong on tissue texture, specular reflections, or instruments, the refinement blends that error into the final disparity.","fun_headline_variants_meta":{"raw":{"variants":["Monocular depth fixes occluded stereo in laparoscopy","Depth-guided refiner cuts laparoscopic disparity error","Semi-supervised refiner uses depth to see past occlusions","Occlusion-aware disparity via monocular depth, no dense labels","DGORNet: depth guidance beats stereo baselines in surgery"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000724,"raw_usage":{"total_tokens":3263,"prompt_tokens":976,"completion_tokens":2287,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":2205}},"tokens_in":592,"tokens_out":2287,"duration_ms":17132,"temperature":1.0,"reasoning_tokens":2205,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:01:50.316115+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle the claim by evaluating DGORNet on a laparoscopic sequence with dense ground truth and computing EPE separately inside and outside occlusion masks: if the depth-guided refinement does not reduce EPE inside occlusion masks compared with the coarse disparity, the central premise is not carrying the gain. A stricter test is to corrupt the monocular depth with a known non-affine bias—such as a depth-dependent drift—and observe whether the final EPE degrades, since Eq. (1) can absorb only scale and shift errors.","supporting_citations":[{"cited_title":"IEEE transactions on medical imaging41(11), 3218–3230 (2022)","cited_arxiv_id":null,"evidence_quote":"Serves as the MSDESIS disparity predictor used in the small prototype and as the source of the ConvBnLRelu building block used in the OPM."},{"cited_title":"In: International Conference on Learning Representations 2021 (2021), https: //openreview.net/forum?id=YicbFdNTTy","cited_arxiv_id":null,"evidence_quote":"Supplies the sinusoidal and cosine position-encoding idea that the Position Embedding module adapts for spatial context."}],"review_version":1}