{"id":"bdff5d7a-e92b-4942-9fdd-3ecf4876f7ad","arxiv_id":"2506.00324","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Combining a difficulty-balancing loss with an occlusion-avoiding loss yields modest but consistent accuracy gains on optical flow and stereo depth benchmarks.","lead":"This paper proposes two loss-weighting schemes, Difficulty Balancing (DB) and Occlusion Avoiding (OA), and their combination, to improve optical flow and stereo depth estimation during training. If the reported gains generalize, the recipe offers a simple, model-agnostic way to boost accuracy on standard benchmarks.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Modest single-run benchmark gains are the core evidence, but no variance estimation is reported and the DB hyperparameter text is internally inconsistent, so the claimed significant improvement is not yet established.","rationale":"The claim stands or falls on whether the reported benchmark deltas are real effects rather than noise. The paper retrains baselines in its own framework and provides ablations over alpha and beta, which is helpful, but every number is a single run. For RAFT on Sintel clean, the best combination improves EPE by 0.08 (1.43 to 1.35) and on KITTI by 0.54; without error bars we cannot tell whether these are systematic. Because the same tables were used to select the combination rule and hyperparameters, the chosen configurations are optimistically biased; a model selected on a benchmark will, on average, look better on that benchmark than its true performance. The internal DB hyperparameter conflict (Section 3.1 and Table 1 versus Section 5.4) further blocks reproduction. This does not make the method wrong; the mechanism is reasonable and the gains are directionally consistent, but it does mean the headline 'significant performance improvements' is premature. A multi-seed replication with confidence intervals is a standard and feasible check, so conditional acceptance, matching the reader, is the right disposition.","tokens_in":12582,"tokens_out":11035,"duration_ms":101181,"concrete_test":"Retrain baseline RAFT, FlowFormer, and RAFT-Stereo plus the proposed combination (multiplication for flow, mask-sum for stereo) with at least 5 random seeds each, using the exact protocol of Sections 4.1–4.2, and report mean plus/minus std for every metric in Tables 1–4. Compute paired per-benchmark differences; if the 95% confidence interval for the improvement includes zero on Sintel clean/final or KITTI, then the claimed significant gain is not established. In the same release, state unambiguously which DB hyperparameters, (2.0, 0.5) or (2.0, 1.0), were used for each main table and rerun Table 6's two disputed cells to resolve the inconsistency.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the DB+OA training losses significantly improve RAFT, FlowFormer, and RAFT-Stereo. The evidence is a set of single training runs per configuration. Reported gains are 0.01–0.1 EPE on Sintel and about 0.5 EPE / 1.9 Fl-all on KITTI; without an estimate of run-to-run variance, these deltas are within the range of normal training noise for recurrent optical-flow models and cannot be separated from seed/optimization stochasticity. The problem is compounded by selection on the evaluation benchmarks: Tables 6–8 are used both to choose alpha, beta, and the combination form (multiplication for flow, mask-sum for stereo) and to support the improvement claim, so the reported numbers are selected maxima and are biased upward. Reproducibility is further weakened by an internal inconsistency: Section 3.1 specifies DB hyperparameters (alpha=2.0, beta=0.5) for optical flow, and the DB row of Table 1 exactly matches the (2.0, 0.5) entry in Table 6, yet Section 5.4 states that '(alpha=2.0, beta=1.0) shows the best overall accuracy.' Consequently, it is unclear which configuration produced the headline results, and Table 2 also shows the finetuned model losing on Sintel clean EPE-all (1.609 to 1.685) while the narrative emphasizes matched/unmatched splits. The improvements are plausible, but the claim of 'significant performance improvements' is not yet supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes two training losses for optical flow and stereo depth estimation: a Difficulty Balancing (DB) loss that reweights the per-pixel L1 loss by an error-based confidence map, and an Occlusion Avoiding (OA) loss that reweights by a forward-backward cycle-consistency confidence map. Several combinations of the two losses are tested, and the authors select multiplicative combination for optical flow and mask-sum for stereo. The losses are applied to RAFT, FlowFormer, and RAFT-Stereo and evaluated on Sintel, KITTI, ETH3D, and Middlebury, with reported improvements over the baselines on most metrics.","tokens_in":12890,"tokens_out":4020,"duration_ms":38378,"significance":"If the reported improvements are reproducible, the contribution is practically useful: it provides simple, model-agnostic training losses that can be plugged into existing recurrent optical flow and stereo architectures. The paper is thorough in its ablation coverage, including hyperparameter grids for both losses and four combination rules across two tasks and several benchmarks. The main weakness is that the central evidence consists of single training runs with differences of 0.01-0.1 EPE on Sintel, which is within the range of typical run-to-run variance for these models. The significance of the claimed improvement is therefore conditional on additional variance analysis and a clear statement of which configuration produced each reported number.","major_comments":[{"comment":"The headline optical-flow gains are small (e.g., 1.43 to 1.35 EPE on Sintel clean, 2.71 to 2.65 on Sintel final, 5.04 to 4.50 on KITTI EPE), yet every configuration is reported from a single training run with no error bars or multiple seeds. For recurrent models trained with stochastic optimization, these deltas are within the range of run-to-run variation, so the Abstract and Section 6 claim of 'significant performance improvements' is not supported by the reported evidence. Please provide multiple seeds per configuration, or a paired statistical test, to establish that the gains exceed training noise. The same concern applies to the stereo results in Table 3.","section":"Section 4.2, Tables 1 and 5"},{"comment":"The DB hyperparameter specification is internally inconsistent. Section 3.1 sets (alpha, beta) = (2.0, 0.5) for optical flow and points to Tables 6 and 8; Table 1's DB row exactly matches the (2.0, 0.5) entry in Table 6 (1.41 / 2.68 / 4.65 / 15.92). However, Section 5.4 states that '(alpha = 2.0, beta = 1.0) shows the best overall accuracy' and that this configuration improves over the RFL loss. The reader cannot determine which configuration produced the headline DB and combination results. Please correct the inconsistency and state the exact loss configuration used for every reported result.","section":"Sections 3.1 and 5.4"},{"comment":"The hyperparameters alpha and beta, as well as the combination rule (multiplication for optical flow, mask-sum for stereo), are selected by evaluating on the same Sintel/KITTI/ETH3D/Middlebury splits that are then used to claim improvement in Tables 1-4. Because the reported numbers are selected maxima over a grid, they are optimistically biased and cannot serve as independent evidence of improvement. Please either fix the configuration before evaluation on a held-out split or report the full grid and clearly describe the selection procedure in the text.","section":"Section 5.4, Tables 6-8"},{"comment":"The finetuned model on Sintel test shows an overall EPE-all regression on the clean pass (1.609 to 1.685), with unmatched EPE worsening from 9.647 to 10.367. The text emphasizes matched-pixel improvements and describes 'significant improvement in the matching area,' but it does not explain or contextualize the overall clean-pass regression. This is directly relevant to the claim of consistent improvement and should be addressed in the discussion.","section":"Table 2"}],"minor_comments":[{"comment":"The setup paragraph contains a typo: 'Florformer' should be 'FlowFormer'.","section":"Section 4.1"},{"comment":"In the description of RAFT-stereo, 'achieves' is misspelled as 'achieves'; please correct the typo.","section":"Section 2.1"},{"comment":"The dataset name is misspelled as 'MiddlueBury'; it should be 'Middlebury'.","section":"Table 3 header"},{"comment":"The OA ablation text says 'We apply five different αs ... to RAFT-Stereo model,' but the table reports optical flow results on RAFT; this should say 'RAFT model.' Also, 'TableTable 8' should be 'Table 8.'","section":"Section 5.4"},{"comment":"The OA loss for stereo requires an additional reverse-disparity pass (the transformation-and-restoration technique in Figure 3), but the training-time computational overhead of this extra pass is not quantified. Please state the additional cost so practitioners can evaluate the trade-off.","section":"Section 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is largely an incremental extension of the authors' own prior work (SCIFlow's regression focal loss and DistractFlow's cycle-consistency confidence); the new element is the combination and hyperparameter search. This is not disqualifying, but the manuscript's internal inconsistency in Section 5.4 and the absence of variance estimates are the main risks. If the authors can provide multi-seed results and clarify the configuration, the contribution would be acceptable for a journal publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe thing to know: this paper is a training-recipe paper, not a model paper. It takes the error-based confidence weighting from SCIFlow (their own prior work) and the cycle-consistency confidence from DistractFlow (also theirs), repackages them as DB and OA losses, and shows that a particular combination (multiplicative for flow, mask-sum for stereo) nudges RAFT, FlowFormer, and RAFT-Stereo down by a few hundredths of EPE on Sintel and a few tenths on KITTI. The components are known; the combination and the stereo extension are new. I agree with the reader that the central claim is plausible, but I do not think the evidence establishes it.\n\nWhat the paper does well: it is honest about provenance, the ablations in Tables 6-8 are reasonably thorough, and the analysis in Table 5 (split by motion magnitude and outlier thresholds) is more informative than most papers at this level. The practical detail of computing right-to-left disparity via transformation-and-restoration for stereo is a nice trick. If the recipe works, it is simple and model-agnostic, which is genuinely useful.\n\nThe soft spots are real. All results are single runs, no error bars, no seeds. The gains are often in the 0.01-0.1 EPE range, which for recurrent models like RAFT is well within normal run-to-run variance. The hyperparameters and combination rules are selected on the same evaluation benchmarks used for the headline numbers, so the reported improvements are selected maxima. Worse, the paper contradicts itself: Section 3.1 says the DB loss for optical flow uses (α=2.0, β=0.5), and Table 1's DB row matches that. But Section 5.4 says (α=2.0, β=1.0) shows the best overall accuracy. It is unclear which configuration produced the reported results. On the test side, the finetuned model is worse on Sintel clean EPE-all (1.609 to 1.685) while the narrative emphasizes matched regions; that is a selective reading. These issues do not kill the paper, but they mean the \"significant performance improvements\" in the abstract are not yet supported.\n\nBottom line: this is a solid, subfield-level contribution that deserves a serious referee. A revision with multi-seed variance estimates, a corrected hyperparameter section, and code release could make it convincing. I would bring it to a reading group for discussion, but I would not cite it as evidence until the variance issue is addressed.","headline":"A plausible but single-run training recipe whose modest gains are not yet separated from noise; worth a careful review but not citable as established.","tokens_in":13407,"tokens_out":4181,"would_cite":false,"duration_ms":34192,"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":"This paper claims that weighting each pixel's training loss by its difficulty—more weight for hard-to-match pixels, less for occluded ones—improves optical flow and stereo depth accuracy on standard benchmarks.","keywords":["optical flow","stereo depth estimation","confidence map","uncertainty","occlusion","cycle consistency","training loss","RAFT"],"falsifier":"Retrain each baseline and each proposed loss at least five times from different random seeds and compare average end-point error with confidence intervals; if the intervals overlap on most benchmarks, the claimed improvements are not established.","tokens_in":12377,"feed_emoji":"🎯","tokens_out":6468,"duration_ms":56305,"temperature":0.7,"pith_summary":"Most optical flow and stereo depth networks are trained with the same loss applied to every pixel, even though some pixels are far harder to learn than others. This paper proposes two training losses that exploit that variation: a Difficulty Balancing (DB) loss, which gives extra weight to pixels where the network's current prediction is far from the ground truth, and an Occlusion Avoiding (OA) loss, which gives less weight to regions that fail a forward-backward consistency check and are therefore likely occluded. Combining the two losses, multiplicatively for optical flow and with a masked sum for stereo, is reported to improve accuracy over the RAFT, FlowFormer, and RAFT-Stereo baselines on Sintel, KITTI, ETH3D, and Middlebury. A reader should care because the change is confined to the loss, leaving the network architecture untouched, so the recipe could transfer to any matching model.","feed_headline":"Per-pixel uncertainty losses sharpen flow and stereo estimates","feed_subtitle":"Two new loss terms, one for hard pixels and one for occlusions, beat RAFT, FlowFormer, and RAFT-Stereo baselines.","key_machinery":"The central objects are two confidence maps. $M_{DB}$ is computed as $\\exp(-\\|f_{gt}-f\\|_2^2)$ from the predicted and ground-truth flow (Eq. 3), marking pixels the model already fits well with values near 1. $M_{OA}$ is computed from forward-backward consistency (Eq. 6), marking regions where matching is reliable and likely non-occluded. The DB loss weights each pixel's L1 error by $1+\\alpha(1-M_{DB})^{\\beta}$, so hard, low-confidence pixels get more gradient; the OA loss weights by $1+\\alpha M_{OA}^{\\beta}$, so occluded, inconsistent pixels stay near their standard weight while confident pixels get more. For stereo, the reverse right-to-left disparity needed for $M_{OA}$ is obtained by flipping the stereo pair, estimating disparity, and flipping back, so the same network supplies both directions. Combining the two maps multiplicatively (optical flow) or with a hard mask plus sum (stereo) prevents the two opposite weighting tendencies from canceling in occluded regions.","core_discovery":"The paper's central claim is that pixel-wise learning difficulty in dense correspondence tasks has two separable causes, and that a loss treating them separately outperforms a uniform loss. The first cause is low model confidence: pixels whose predicted flow or disparity is far from the ground truth are harder, so the DB loss multiplies their L1 error by a factor $1+\\alpha(1-M_{DB})^{\\beta}$, where $M_{DB}=\\exp(-\\|f_{gt}-f\\|_2^2)$ is an error-based confidence map. The second cause is occlusion: pixels without a real correspondence should not be forced to match, so the OA loss uses a cycle-consistency map $M_{OA}$ from Eq. (6) and weights the error by $1+\\alpha M_{OA}^{\\beta}$, keeping occluded regions near the standard L1 weight while boosting confident regions. The paper reports that the multiplicative combination is best for optical flow, the mask-sum combination is best for stereo, and that the combined loss improves over all three baselines.","pith_inferences":["Because the losses change only the training objective, they could be dropped into other dense matching networks, such as scene flow or video depth models, where hard pixels and occlusions also vary spatially; the paper does not test this.","The same cycle-consistency map used for OA could serve as a self-supervised weight in settings without ground truth, making the recipe applicable to semi-supervised or unsupervised training; this is an extension, not a claim of the paper.","The hyperparameters $\\alpha$ and $\\beta$ are chosen by validation grid search; an annealed schedule over training, where the focus shifts from hard pixels to occlusion avoidance, is a natural variation the authors do not explore.","Since the benchmarks are single runs, a practical adopter would first replicate the comparison over multiple seeds before deciding; the paper itself reports no variance statistics."],"forward_implications":["Applied to RAFT, the combined loss reduces Sintel clean EPE from 1.43 to 1.35 and KITTI Fl-all from 17.4 to 15.45, according to the paper's Table 1.","Applied to FlowFormer, the multiplicative combination lowers Sintel clean EPE from 1.01 to 0.97 and KITTI Fl-all from 14.72 to 14.17.","Applied to RAFT-Stereo, the mask-sum combination lowers ETH3D bad-1.0 from 2.44 to 1.67 and KITTI D1-all from 1.96 to 1.83 on the reported test sets.","The DB loss mainly helps large-displacement and object-boundary pixels, while OA mainly helps small-displacement and occluded regions, as separated in Table 5."],"supporting_citations":[{"why":"Supplies the RAFT optical flow baseline that the DB and OA losses are applied to and compared against.","marker":"[28]"},{"why":"Supplies the FlowFormer transformer baseline that the combined multiplicative loss is shown to improve.","marker":"[5]"},{"why":"Supplies the RAFT-Stereo baseline for stereo depth experiments.","marker":"[19]"},{"why":"Introduces the Regression Focal Loss that the DB loss extends with new hyperparameters and a stereo version.","marker":"[18]"},{"why":"Provides the error-based confidence map formula that defines M_DB in Eq. (3).","marker":"[6]"},{"why":"Supplies the cycle-consistency confidence map (Eq. 6) that the OA loss is built on.","marker":"[7]"},{"why":"Provides the forward-backward consistency check used to derive occlusion regions.","marker":"[22]"},{"why":"Supplies the FlyingThings3D training data used for optical flow and stereo pretraining.","marker":"[21]"}],"fun_headline_variants":["Uncertainty-weighted losses fix flow and stereo errors","Pixel-wise difficulty losses boost flow and stereo","Two losses: one for hard pixels, one for occlusions","Uncertainty maps steer training away from occluded pixels","Occlusion-aware loss plus difficulty weight improves flow"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains come from single training runs with no error bars, so the paper assumes that differences of about 0.01 to 0.1 pixels in average error (for example, 1.43 vs 1.35 on Sintel clean) are real improvements rather than run-to-run training noise.","fun_headline_variants_meta":{"raw":{"variants":["Uncertainty-weighted losses fix flow and stereo errors","Pixel-wise difficulty losses boost flow and stereo","Two losses: one for hard pixels, one for occlusions","Uncertainty maps steer training away from occluded pixels","Occlusion-aware loss plus difficulty weight improves flow"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000405,"raw_usage":{"total_tokens":2107,"prompt_tokens":948,"completion_tokens":1159,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":564,"completion_tokens_details":{"reasoning_tokens":1082}},"tokens_in":564,"tokens_out":1159,"duration_ms":8416,"temperature":1.0,"reasoning_tokens":1082,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:07:13.456157+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain each baseline and each proposed loss at least five times from different random seeds and compare average end-point error with confidence intervals; if the intervals overlap on most benchmarks, the claimed improvements are not established.","supporting_citations":[{"cited_title":"Raft: Recurrent all-pairs field transforms for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the RAFT optical flow baseline that the DB and OA losses are applied to and compared against."},{"cited_title":"Flowformer: A transformer architecture for optical flow","cited_arxiv_id":null,"evidence_quote":"Supplies the FlowFormer transformer baseline that the combined multiplicative loss is shown to improve."},{"cited_title":"Raft-stereo: Multilevel recurrent field transforms for stereo matching","cited_arxiv_id":null,"evidence_quote":"Supplies the RAFT-Stereo baseline for stereo depth experiments."},{"cited_title":"Sciflow: Empowering lightweight optical flow models with self-cleaning iterations","cited_arxiv_id":null,"evidence_quote":"Introduces the Regression Focal Loss that the DB loss extends with new hyperparameters and a stereo version."},{"cited_title":"Liteflownet3: Resolv- ing correspondence ambiguity for more accurate optical flow estimation","cited_arxiv_id":null,"evidence_quote":"Provides the error-based confidence map formula that defines M_DB in Eq. (3)."},{"cited_title":"Distractflow: Improving optical flow estimation via real- istic distractions and pseudo-labeling","cited_arxiv_id":null,"evidence_quote":"Supplies the cycle-consistency confidence map (Eq. 6) that the OA loss is built on."},{"cited_title":"Unflow: Un- supervised learning of optical flow with a bidirectional cen- 9 sus loss","cited_arxiv_id":null,"evidence_quote":"Provides the forward-backward consistency check used to derive occlusion regions."}],"review_version":1}