{"id":"17c3d61b-e00a-4864-8737-b7cfd2a0e45f","arxiv_id":"1908.05898","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"OFNet shares a common occlusion cue across two specialized feature paths and adds context-aware modules, achieving state-of-the-art occlusion edge and orientation prediction on PIOD and BSDS ownership.","lead":"This paper introduces OFNet, a neural network that predicts which object occludes another in an image. It reports improved accuracy over previous methods on two standard occlusion reasoning datasets.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA gains in Tables 1-2 may reflect test-set hyperparameter selection: fusion ratio (Table 5) and stripe kernel (Table 6) are chosen from the same test set with no validation split or error bars.","rationale":"I read the paper in good faith. The architecture is coherent: sharing an occlusion cue while separating high-level features is a reasonable design, and the release of source code is a real independent artifact that could support reproduction. The ablations in Tables 3-6 are internally consistent, and the qualitative examples are plausible. The single weakest link is not the architecture itself but the evaluation protocol: every design choice--fusion ratio, stripe kernel, even the dilated rates--is selected by inspecting performance on the same test sets used for the headline numbers, and the paper reports no variance. Since the claimed margins over DOOBNet are not large relative to the spread among tested variants, the reported improvements could be inflated by selection on the test set. This is exactly the reader's weakest assumption, and I agree with it. It does not warrant rejection; the paper's conclusion is conditionally acceptable. The check I propose--validation-based selection and repeated-seed error bars--would settle whether the concern lands. Until then, UNCHANGED (CONDITIONAL) is the right verdict.","tokens_in":12079,"tokens_out":7426,"duration_ms":72477,"concrete_test":"Retrain the released OFNet and the DOOBNet baseline on PIOD under identical settings with at least five random seeds; report mean and standard deviation of AP for the final 64:16 / 3x11 configuration. In parallel, perform model selection on a held-out validation split (e.g., 20% of the training set, not the test set) over the same fusion-ratio and kernel-size grids, then evaluate the selected configuration once on the official test set. If the test AP of the validation-selected model is more than about 1-2 points below the reported 0.729 AP, or if the DOOBNet gap falls within the seed-to-seed noise, the SOTA claim should be downgraded to a more cautious statement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim--\"remarkably advances state-of-the-art\" with +4.6 AP on PIOD and +6.1 AP on BSDS (Sec. 4.2, Tables 1-2)--rests on the final configuration's numbers being trustworthy point estimates. The paper never fixes these numbers before seeing the test labels. In Sec. 4.3, Table 5 selects fusion ratio 64:16 by comparing five ratios on the PIOD test set; Table 6 selects 3x11 stripe convolution by comparing five kernel sizes on the same test set. No validation split is used, no random seeds are repeated, and no error bars are reported. The selected-vs-adjacent gaps are comparable to the headline improvement: 64:16 vs 80:16 changes AP by 0.007 (0.729 vs 0.722) and 3x11 vs 3x9 changes AP by 0.017 (0.729 vs 0.712), while the claimed gain over DOOBNet is 0.046. On BSDS, with only 100 test images, AP estimates have even higher variance. Thus, the reported SOTA numbers are the maximum of a small grid of models evaluated on the benchmark that is also used for model selection, so part of the advantage may be optimistic selection bias rather than the architecture itself. This is load-bearing because the main quantitative evidence for the paper's thesis depends on it.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes OFNet, a CNN for occlusion relationship reasoning that jointly outputs an occlusion edge map and per-pixel occlusion orientation. The architecture uses an encoder-decoder backbone with two paths: an edge path that aggregates low- and high-level cues, and an orientation path that extracts a new \"bilateral feature\" via a Multi-rate Context Learner (MCL), fuses it with the shared occlusion cue in a Bilateral Response Fusion (BRF) module, and reasons with orthogonal stripe convolutions. On the PIOD and BSDS ownership datasets, the authors report state-of-the-art results, exceeding the DOOBNet baseline by 4.6% and 6.1% AP respectively on the occlusion-relationship metric, and they provide ablations of the main components (Tables 3-6). The source code is released.","tokens_in":12354,"tokens_out":9654,"duration_ms":82657,"significance":"The paper makes a genuinely useful architectural contribution to occlusion relationship reasoning: the explicit separation of the edge and orientation paths with a shared occlusion cue addresses a design tension that single-stream methods (DOOBNet) and fully separate networks (DOC) did not resolve, and the MCL, BRF, and stripe-convolution modules are well motivated by the geometry of the task. The ablations in Tables 3-6 are systematic, and the monotonic trends (e.g., orientation AP rising from 0.703 to 0.729 as the bilateral-feature ratio increases in Table 5) give qualitative support for the design choices. The release of source code is a further strength. The central quantitative claim -- 4.6% AP gain on PIOD and 6.1% AP on BSDS ownership over DOOBNet -- is plausible but, as detailed in the major comments, its magnitude is not yet rigorously supported because of test-set hyperparameter selection, the absence of error bars, and the small BSDS test set.","major_comments":[{"comment":"The two key hyperparameters of the final model, the 64:16 fusion ratio in BRF and the 3x11 stripe kernel, are selected by comparing five variants on the PIOD test set (Tables 5 and 6), and the same test set is then used for the headline comparisons in Tables 1 and 2; no validation split, repeated-seed runs, or error bars are reported anywhere in the paper. The size of the selection effect is comparable to the headline gain: the 64:16 vs 80:16 difference is 0.007 AP (0.729 vs 0.722) and the 3x11 vs 3x9 difference is 0.017 AP (0.729 vs 0.712), against a reported 0.046 AP improvement over DOOBNet. The reported numbers are therefore the maximum of a small grid of configurations evaluated on the benchmark that was also used for model selection, so part of the claimed margin may reflect optimistic selection bias rather than the architecture itself. The authors should either choose hyperparameters on a held-out validation split and then report the corresponding test numbers, or report the mean and standard deviation over multiple random seeds for the final configuration and for DOOBNet, so that the magnitude of the improvement can be assessed.","section":"Sec. 4.3 (Tables 5 and 6); Sec. 4.2 (Tables 1 and 2)"},{"comment":"The BSDS ownership dataset contains only 100 test images, and the claimed 6.1% AP gain over DOOBNet (0.440 to 0.501) is a point estimate from a single run. On a test set of this size, ODS/OIS/AP differences of a few points are within plausible sampling variation, and the paper gives no confidence intervals, bootstrap estimates, or per-image paired statistics. As it stands, the BSDS half of the \"remarkably advances state-of-the-art\" claim is not statistically assessable; the authors should add some variance estimate or a paired comparison over the 100 test images.","section":"Sec. 4.2 (Table 1, BSDS ownership)"}],"minor_comments":[{"comment":"Equation (3) contains a notation error: \"SL(f(ai,a i))\" has an undefined function f and duplicated arguments, and the value of the loss weight lambda is never given in Sec. 4.1 despite being a free parameter of the method; the authors should provide a correct expression (including, if used, the angular wrap-around in the smooth L1 term) and report lambda, along with the learning-rate schedule, batch size M, and number of iterations, which are also absent from Sec. 4.1.","section":"Sec. 3.3, Eq. (3); Sec. 4.1"},{"comment":"In the conclusion, the proposed module is referred to as \"an MSL is proposed\" while everywhere else it is called MCL; this typo should be corrected.","section":"Sec. 5 (Conclusion)"},{"comment":"In Table 3, the right-hand OPR columns of \"Baseline(split decoder)\" are numerically identical to those of \"Baseline\" (ODS .702, OIS .712, AP .683); if the split decoder leaves the orientation output unchanged by construction, this should be stated explicitly, otherwise the row looks like a copy-and-paste error.","section":"Table 3"},{"comment":"Table 7 shows that the 11x11 kernel slightly outperforms the selected 3x11 kernel (e.g., PIOD OPR AP 0.732 vs 0.729), and the main text should state this trade-off explicitly rather than only in the appendix, since it bears on the claim that the stripe kernel is the preferred design.","section":"Appendix A.4 (Table 7)"},{"comment":"Figure 3 is difficult to read: the block labels (res1 through res5, decoder output, BRF output, upsampling factors) are typeset in very small font, and the caption does not explain the meaning of different block widths and heights; a redrawn figure with a legend would substantially improve the clarity of the architecture description.","section":"Fig. 3"},{"comment":"The comparison set contains only methods published between 2015 and 2018 (SRF-OCC, DOC-HED, DOC-DMLFOV, DOOBNet); adding concurrent or more recent occlusion-reasoning methods to the comparison would make the \"state-of-the-art\" claim more robust.","section":"Sec. 4.2 (Tables 1 and 2)"}],"recommendation":"major_revision","confidential_remarks":"The absence of error bars and the selection of hyperparameters on the test set are common in this subfield, and the monotonic ablation trends suggest the qualitative conclusions would survive a more careful protocol; however, the paper's headline is explicitly quantitative, so for a journal version the authors should be required to add variance estimates or a validation-based selection procedure. I also note that the edge path is largely inherited from the authors' own ICME work [15]; the novelty relative to [15] (the added refinement block, as stated in Sec. 3.1) should be spelled out in a revised version so that the contribution boundary is unambiguous."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a reasonable architecture paper for occlusion reasoning, but the headline numbers are a bit less robust than they look. The central design—edge and orientation paths that share the decoder's occlusion cue while keeping high-level features separate—is a legitimate response to the relevance/distinction tension in the task. MCL is close to ASPP plus a parallel 1x1 conv, stripe convolution is a factorized large-kernel trick, and the edge path builds on the authors' prior work; but the specific combination is new and the paper demonstrates its value with careful ablations. Tables 3-6 and the appendix cover branch splitting, feature removal, fusion ratios, and kernel shapes; the code is public. That is real credit.\n\nThe soft spots are mainly methodological and they matter for the exact magnitude of the claim. Tables 5 and 6 select the fusion ratio (64:16) and stripe kernel (3x11) by comparing variants on the PIOD test set, with no validation split and no error bars. The gaps between the chosen and adjacent variants (0.007 AP for the ratio, 0.017  for the kernel) are not negligible next to the claimed 0.046 AP gain over DOOBNet. On BSDS, with only 100 test images, variance is even higher. So the reported SOTA numbers are essentially the best of a small grid, and part of the advantage could be selection bias. That said, the improvements over DOOBNet are consistent across ODS, OIS, and AP on both datasets, so it would be surprising if the architecture contributed nothing. The prose's \"remarkably advances\" is stronger than the evidence warrants, but the core result is plausible.\n\nWho is this for? The occlusion relationship / contour detection community, where it is a modest but legitimate step. A serious referee should engage with it, but should ask for multiple runs or a validation-based hyperparameter selection before the SOTA claim is taken at face value. I would not cite it in my own work unless I were working on this exact task, and I would not bring it to a general reading group, but it deserves to be reviewed rather than desk-rejected.","headline":"The OFNet two-path architecture is a sensible, well-ablated incremental contribution, but the headline SOTA gains are less solid than the prose suggests because the key hyperparameters were selected on the test set.","tokens_in":12913,"tokens_out":1874,"would_cite":false,"duration_ms":18904,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Sharing the occlusion cue while separating semantic feature paths improves occlusion relationship reasoning on PIOD and BSDS ownership datasets.","keywords":["occlusion boundary detection","occlusion orientation estimation","border ownership","bilateral feature","multi-rate context learner","stripe convolution","PIOD dataset","BSDS ownership"],"falsifier":"Retrain OFNet and DOOBNet from the same initializations, tune the fusion ratio and stripe size on a held-out portion of the training data, then evaluate on the untouched test split; if the 64:16 ratio and 3x11 kernel no longer beat smaller variants by comparable margins, the claimed advantage is benchmark-specific.","tokens_in":11854,"feed_emoji":"🧩","tokens_out":5254,"duration_ms":48070,"temperature":0.7,"pith_summary":"This paper argues that occlusion relationship reasoning—deciding which object is in front and where the boundary lies—needs a network that treats occlusion edge and occlusion orientation as neither fully shared nor fully separate tasks. The authors propose OFNet, which shares the occlusion cue across the two prediction paths but learns high-level semantic features in two separate streams, and adds a bilateral feature that captures context from both the foreground and background sides of a contour. This design targets two shortcomings in prior CNN-based methods: losing the shared relevance between edge and orientation by using separate networks, and losing orientation-specific context by using a single shared stream. On the PIOD and BSDS ownership datasets, OFNet outperforms the DOOBNet baseline by 4.6% and 6.1% average precision for orientation prediction. If the result holds, the practical message is that occlusion reasoning improves when the network architecture mirrors the task's shared-and-separate structure.","feed_headline":"Network beats occlusion-relation baselines by up to 6.1 points","feed_subtitle":"Sharing the occlusion cue but splitting semantic feature paths lifts PIOD and BSDS ownership results.","key_machinery":"The load-bearing machinery is the bilateral feature, defined as the contextual information on both sides of an occlusion contour that indicates which side is foreground. The Multi-rate Context Learner (MCL) extracts it by applying dilated convolutions with rates 6, 12 and 18 together with a local 1x1 convolution branch, then fuses the branches with another 1x1 convolution; this gives an edge pixel a receptive field large enough to see both occluding and occluded objects. The Bilateral Response Fusion (BRF) module concatenates this bilateral feature with the shared occlusion cue from the decoder and passes the pair through two 3x3 convolutions, producing a fused orientation map. Finally, the Occlusion Relationship Reasoning module replaces a single 3x3 kernel with orthogonal stripe convolutions (11x3 and 3x11), which correspond to decomposing a contour's tangent direction into vertical and horizontal components, so the network perceives scene context on both sides of the edge without the memory cost of a full 11x11 kernel. Together these modules implement the paper's central design rule: share the location cue, separate the semantic features.","core_discovery":"OFNet claims that a single-stream shared decoder is the wrong abstraction for occlusion relationship reasoning. The network keeps one occlusion cue—the decoder feature that locates the boundary—and feeds it to both the edge path and the orientation path, while the high-level features themselves are split so each path learns the semantics it needs. For orientation, OFNet introduces the bilateral feature: a Multi-rate Context Learner with parallel dilated convolutions (rates 6, 12, 18) plus a 1x1 local branch reads foreground and background regions at multiple scales, and a Bilateral Response Fusion module combines that bilateral feature with the occlusion cue at a 64:16 channel ratio. A pair of orthogonal stripe convolutions (11x3 and 3x11) then aggregates surrounding scene context along the contour's vertical and horizontal tendencies. The result is an occlusion edge map and orientation map whose joint accuracy exceeds prior systems; on PIOD the orientation AP is 0.773 versus 0.723 for DOOBNet, and on BSDS ownership it is 0.585 versus 0.501.","pith_inferences":["The paper does not test learned or adaptive fusion ratios; a gating mechanism that lets the network decide how much bilateral context to trust per pixel could be tested against the fixed 64:16 ratio.","The share-what-is-common, separate-what-is-specific principle could generalize to other joint prediction tasks, such as simultaneous semantic edge detection and instance segmentation, where boundary location and object identity have a similar structure.","The MCL's explicit local 1x1 branch alongside dilated rates implies that pure multi-scale context loses nearby contour cues; a similar local-global split could benefit other boundary-based tasks such as border ownership from stereo cues.","One could probe whether the 3x11 stripe advantage is tied to the image statistics of PIOD and BSDS by evaluating on rotated or scaled versions of the same images."],"forward_implications":["Occlusion reasoning in natural images improves when the edge and orientation paths share only the boundary-location cue rather than all high-level features.","Orientation prediction benefits from explicit bilateral context, which can be learned by multi-rate dilated convolution and fused with the occlusion cue.","Stripe convolution gives a larger effective receptive field along contour directions at modest computational cost, replacing plain 3x3 convolution in this task.","The architecture transfers to the small-data BSDS ownership dataset, gaining over DOOBNet, which suggests the design also helps when training data are scarce."],"supporting_citations":[{"why":"Defines the occlusion edge and orientation representation and provides the PIOD dataset used for training and evaluation.","marker":"[21]"},{"why":"DOOBNet is the single-stream shared-decoder baseline whose design is contrasted and whose scores are surpassed.","marker":"[31]"},{"why":"Supplies the contour-extraction structure adopted and adapted for the edge path.","marker":"[15]"},{"why":"Provides the BSDS ownership dataset and the figure/ground assignment task on which the method is evaluated.","marker":"[22]"},{"why":"The ASPP multi-rate dilated convolution design that the MCL extends and explicitly differentiates itself from.","marker":"[3]"},{"why":"SRF-OCC is an earlier border-ownership baseline included in the comparisons.","marker":"[30]"},{"why":"Used for adjusting the predicted orientation to the tangent direction of the final occlusion edge map.","marker":"[11]"},{"why":"Provides the ResNet-50 backbone from which the network is fine-tuned.","marker":"[9]"}],"fun_headline_variants":["Shared cue, split paths: occlusion reasoning up 6.1 points","OFNet's dual-stream trick lifts PIOD orientation AP to 0.773","One cue, two feature paths: new best in occlusion reasoning","Bilateral context and stripe convs sharpen occlusion maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported gains are the best of several fusion-ratio and stripe-kernel variants selected on the test set, so the load-bearing premise is that those choices (64:16 fusion ratio, 3x11 stripes) reflect a real advantage rather than overfitting the specific benchmarks.","fun_headline_variants_meta":{"raw":{"variants":["Shared cue, split paths: occlusion reasoning up 6.1 points","OFNet's dual-stream trick lifts PIOD orientation AP to 0.773","One cue, two feature paths: new best in occlusion reasoning","Bilateral context and stripe convs sharpen occlusion maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000468,"raw_usage":{"total_tokens":2349,"prompt_tokens":976,"completion_tokens":1373,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":592,"completion_tokens_details":{"reasoning_tokens":1298}},"tokens_in":592,"tokens_out":1373,"duration_ms":10515,"temperature":1.0,"reasoning_tokens":1298,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T13:01:03.234078+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain OFNet and DOOBNet from the same initializations, tune the fusion ratio and stripe size on a held-out portion of the training data, then evaluate on the untouched test split; if the 64:16 ratio and 3x11 kernel no longer beat smaller variants by comparable margins, the claimed advantage is benchmark-specific.","supporting_citations":[{"cited_title":"Doc: Deep occlusion estimation from a single image","cited_arxiv_id":null,"evidence_quote":"Defines the occlusion edge and orientation representation and provides the PIOD dataset used for training and evaluation."},{"cited_title":"Doobnet: Deep object occlusion boundary detection from an image","cited_arxiv_id":null,"evidence_quote":"DOOBNet is the single-stream shared-decoder baseline whose design is contrasted and whose scores are surpassed."},{"cited_title":"Context-Constrained Accurate Contour Extraction for Occlusion Edge Detection","cited_arxiv_id":"1903.08890","evidence_quote":"Supplies the contour-extraction structure adopted and adapted for the edge path."},{"cited_title":"Fig- ure/ground assignment in natural images","cited_arxiv_id":null,"evidence_quote":"Provides the BSDS ownership dataset and the figure/ground assignment task on which the method is evaluated."},{"cited_title":"Deeplab: Semantic image segmentation with deep convolutional nets, atrous convolu- tion, and fully connected crfs","cited_arxiv_id":null,"evidence_quote":"The ASPP multi-rate dilated convolution design that the MCL extends and explicitly differentiates itself from."},{"cited_title":"Fast 2d border ownership assignment","cited_arxiv_id":null,"evidence_quote":"SRF-OCC is an earlier border-ownership baseline included in the comparisons."},{"cited_title":"An online learning approach to occlusion boundary detection","cited_arxiv_id":null,"evidence_quote":"Used for adjusting the predicted orientation to the tangent direction of the final occlusion edge map."}],"review_version":1}