{"id":"66062cae-f733-4765-aeba-d666b5e19a65","arxiv_id":"2506.21198","paper_version":2,"verdict":"CONDITIONAL","confidence":"LOW","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"UNLOCK adapts a pinhole-trained amodal panoptic segmentation model to unlabeled panoramic images with no source data, via omni pseudo-labeling and amodal-driven object mixing.","lead":"This paper introduces a new task, Source-Free Occlusion-Aware Seamless Segmentation, in which a model pre-trained on labeled pinhole images must adapt to unlabeled panoramic images without access to the original source data. The proposed UNLOCK framework, based on omni pseudo-labeling and amodal-driven object mixing, is reported to match or exceed methods that still use source data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported gains may be inflated by test-set hyperparameter selection: Section 9 sweeps thresholds on K2B with no validation split described.","rationale":"The reader's verdict is CONDITIONAL, and this critique reinforces that conditionality rather than moving it to ACCEPT or REJECT. The most load-bearing weakness is not the OPLL semantic-reliability assumption (Eq. 3), though that is a legitimate robustness concern; it is the validation protocol. The paper's Section 9 hyperparameter analysis appears to select thresholds directly on the 100-image test set of K2B, with no validation split, and the sweeps show that mAPQ varies by more than 1 point across plausible settings—enough to account for a meaningful fraction of the reported +4.3 improvement. Because the central claim is explicitly quantitative ('state-of-the-art scores of 10.9 in mAAP and 11.6 in mAP'), the missing validation discipline directly threatens the headline result. The method itself is reasonable and the new A2B benchmark is a useful contribution; the failure-case discussion (Sec. 10.2) and the ablation tables indicate genuine effort. The paper also honestly discloses limitations (Sec. 11.1). No formal verification or code is provided, so the empirical validation is the only support. Therefore the appropriate verdict remains CONDITIONAL: accept only if the authors provide a validation-based hyperparameter selection and rerun the final evaluation, or show that the reported numbers persist under fixed/default thresholds. The reader's rationale did mention the validation issue, but the formal 'weakest_assumption' field pointed elsewhere; hence partial agreement.","tokens_in":27478,"tokens_out":3229,"duration_ms":39625,"concrete_test":"Ask the authors to report the exact data split and, if any labeled target images were used for hyperparameter selection, to re-run UNLOCK with hyperparameters fixed on a held-out validation set (e.g., a subset of the 2,000 unlabeled target images with pseudo-labels, or a split of the 100 labeled images) and report the final test numbers. Concretely: hold out 20 of the 100 labeled BlendPASS images as validation, select τ_fix, τ_per, τ'_fix, τ'_per, and R on those 20 images only, then evaluate on the remaining 80 test images and report all five metrics. If the mAPQ gain over Source-only drops below the reported +4.3 (or mAAP/mAP fall below 10.9/11.6), the headline 'state-of-the-art' claim is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim—UNLOCK achieves state-of-the-art mAAP of 10.9, mAP of 11.6, and +4.3 mAPQ over source-only—rests on hyperparameters chosen in Section 9. The K2B target domain is described as 2,000 unlabeled training images and 100 labeled test images (Sec. 4.1), and no validation or development split is ever mentioned. Section 9 says thresholds are varied 'on the KITTI360-APS → BlendPASS benchmark,' with the final values taken from the configuration that maximizes mAPQ while 'also taking into account the overall performance' (gray rows in Tables 9–19). Since the 100 labeled test images are the only labeled target data, this procedure effectively selects hyperparameters using test labels. The sweeps show non-trivial sensitivity: Table 18 varies mAPQ from 24.80 to 25.84 across τ'_fix/τ'_per, and Table 19 varies mAPQ from 24.83 to 25.84 across R; the chosen R=10 is the peak. Thus a substantial part of the reported +4.3 mAPQ improvement and the mAAP/mAP SOTA numbers may reflect test-set overfitting rather than the method's intrinsic source-free adaptation ability. This also makes the comparison to UDA baselines unfair if those baselines were not tuned on the same labels. The claim 'comparable to source-dependent methods' is therefore not yet substantiated by the evidence as presented.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript introduces Source-Free Occlusion-Aware Seamless Segmentation (SFOASS), a task setting in which a model trained on labeled pinhole images must adapt to unlabeled panoramic images without access to the source data. The proposed UNLOCK framework combines Omni Pseudo-Labeling Learning (OPLL), which generates class-wise thresholded pseudo-labels for semantic, instance, and amodal instance branches, with Amodal-Driven Context Learning (ADCL), which builds an object pool from high-confidence amodal predictions and pastes objects into training images with zeroed occlusion regions. Experiments are reported on KITTI360-APS→BlendPASS (K2B), a newly introduced AmodalSynthDrive→BlendPASS (A2B) benchmark, and Cityscapes-to-DensePASS (C2D) for semantic segmentation. The paper claims that UNLOCK achieves state-of-the-art mAAP of 10.9 and mAP of 11.6, and an absolute improvement of +4.3 mAPQ over the source-only model, reaching performance comparable to source-dependent UDA methods.","tokens_in":27964,"tokens_out":3946,"duration_ms":49220,"significance":"If the reported results are reliable, the paper would make a useful contribution by extending source-free domain adaptation to the relatively new occlusion-aware seamless segmentation task and by providing a synthetic-to-real benchmark for that task. The task formulation is clear, the method is described in enough detail to be implemented, and the authors commit to releasing data and code. The paper also includes ablations of the two main components and additional results on panoramic semantic segmentation. However, the central quantitative claims are currently undermined by the apparent use of the target test set for hyperparameter selection and by the absence of any uncertainty quantification. The claimed state-of-the-art numbers and the headline +4.3 mAPQ gain therefore cannot yet be taken as evidence for the method's intrinsic source-free adaptation ability; they must be re-established under a clean validation protocol.","major_comments":[{"comment":"The hyperparameter sweeps in Section 9 are performed directly on the K2B benchmark, whose only labeled target data are the 100 test images (Sec. 4.1), and the reported final configurations are selected as the gray rows that maximize mAPQ. Table 18 shows mAPQ ranging from 24.80 to 25.84 as tau'_fix and tau'_per vary, and Table 19 shows mAPQ ranging from 24.83 to 25.84 as R varies, with the chosen R=10 at the peak. This is effectively test-label hyperparameter selection, and it means the reported +4.3 mAPQ improvement and the mAAP/mAP state-of-the-art numbers partly reflect tuning on the test set rather than the method's intrinsic performance. Please introduce a proper validation split (or cross-validation) for threshold selection, or justify fixed thresholds by a source-only or model-free criterion, and report final results on the held-out test set with the chosen configuration.","section":"Section 9, Tables 18-19"},{"comment":"All experimental results are reported as single runs without error bars or multiple seeds. Since the BlendPASS target test set contains only 100 images, the reported margins on instance-level metrics (e.g., mAAP 10.9 vs. 10.5 for the source-only baseline, or mAP 11.6 vs. 11.3 for the EDAPS baseline) are small and could easily lie within run-to-run variation. The paper should report means and standard deviations over at least three seeds, and ideally a significance test, for the main tables and the component ablations, before claims such as 'state-of-the-art' or '+4.3 improvement' are substantiated.","section":"Section 4.3, Tables 1-7"},{"comment":"OPLL revises every instance and amodal instance mask by intersecting it with the semantic branch's Thing mask. This assumes the semantic branch is more reliable than the instance branches on the target panoramic domain, but the paper provides no evidence for this assumption on classes where the semantic branch is weak. For example, in Table 1 the semantic branch achieves mIoU of only 18.3 for cyclists and 0.0 for two-wheelers on the K2B benchmark, so the semantic Thing mask may suppress correct instance predictions for exactly the classes where pseudo-labels are most needed. Please quantify how many instance proposals are discarded by Eq. (3), report per-class agreement between the semantic and instance branches, and add an ablation that removes or softens this semantic masking step.","section":"Section 3.2, Eqs. (2)-(3)"},{"comment":"The comparison to UDA baselines is not yet framed fairly. UNLOCK's thresholds are tuned on the K2B target test set in Section 9, while the reproduced UDA baselines on A2B are stated in Section 8 to use hyperparameters from their original papers. It is not stated whether the same K2B-tuned thresholds are transferred unchanged to the A2B benchmark or whether the baselines were given any comparable validation-based tuning. This asymmetry could inflate the reported advantage over source-dependent methods. Please state the exact hyperparameters used for UNLOCK on A2B, and if they are the K2B-tuned values, either justify the transfer or re-tune all methods on a common validation protocol before computing the final numbers.","section":"Section 4.3 and Section 8"}],"minor_comments":[{"comment":"The text says 'The analysis of these thresholds is provided in the supplementary material,' but the threshold analysis appears in Section 9 of the main paper; this cross-reference should be corrected.","section":"Section 4.2"},{"comment":"The right half of Table 2 labels the Instance Segmentation metric as 'mAAP for Instance Segmentation (IS)'; this should be 'mAP' to match the metric name used in the abstract and Figure 1.","section":"Table 2"},{"comment":"Equation (6) is notationally unclear: the expressions '1_{P_j ...}' appear to use an undefined summation symbol, and the equality mixes set-level and pixel-level notation. The uncertainty region should be defined with explicit spatial indices so that the intersection and the indicator functions are unambiguous.","section":"Section 3.2, Eq. (6)"},{"comment":"The caption of Figure 13 is identical to the caption of Figure 12 and describes a failure case, but the figure actually shows qualitative examples of the amodal-driven object pool; the caption should be rewritten accordingly.","section":"Figure 13"},{"comment":"Several entries in Table 1 contain corrupted formatting with multiple concatenated numbers (e.g., '10.101.3' for DATR and '15.304.241.1' for Trans4PASS under mAPQ); the rendered table should be checked so that each cell contains exactly one number.","section":"Table 1"},{"comment":"The phrase 'comparable results' with respect to UnmaskFormer should be qualified, since Table 1 shows UNLOCK is about one point lower on mPQ and mIoU while being higher on instance metrics; the current wording is stronger than the evidence in the table.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The core issue is the target-test-set hyperparameter selection in Section 9. If the authors can re-run the evaluation with a clean validation protocol, or convincingly argue that the thresholds are not fit to the test labels, the paper could become acceptable. The manuscript also needs multi-seed reporting before the state-of-the-art claims can be taken seriously. I do not see an internal inconsistency that would require rejection, but the current evidence for the headline numbers is insufficient."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is the first paper to formulate source-free adaptation for amodal panoptic segmentation on panoramic images (SFOASS), and it ships a new synthetic-to-real benchmark (AmodalSynthDrive→BlendPASS). That task framing and benchmark are useful even if every individual trick is known. Second, the headline numbers — SOTA mAAP 10.9, mAP 11.6, +4.3 mAPQ over source-only — rest on hyperparameters selected using the target test set, and that is not a minor footnote.\n\nWhat's good: the method is a sensible recombination. Class-wise pseudo-label thresholds, uncertainty-weighted BCE, and object-pool mixing that preserves amodal shapes are established ideas, but the combination is coherent and the ablations show each component contributes. Gains are consistent across two benchmarks and also on the C2D semantic segmentation extension. They include a failure case and a limitations section, which is more than many papers do. Code and data are promised.\n\nWhere it's soft: Section 9 sweeps thresholds on K2B and chooses the configuration that maximizes mAPQ on the 100-image labeled test set; no validation split is described. Tables 18 and 19 show about a 1-point mAPQ swing depending on τ′ and R, and R=10 is the peak of the sweep. So part of the reported +4.3 improvement may be test-set tuning, and the UDA baselines were presumably not given the same privilege. There are also no error bars anywhere, which matters with only 100 test images. The OPLL assumption that the semantic branch is more reliable than instance branches on target data is plausible but unexamined; one ablation that tests this would strengthen the paper.\n\nWho it's for: people working on panoramic perception, amodal segmentation, and source-free adaptation. The paper deserves a serious referee: the task is new, the experiments are extensive, and the flaw is addressable. Recommendation: send to peer review, but ask for a validation split (or a clear hyperparameter-selection protocol that does not use test labels), multiple seeds, and the code artifacts.","headline":"First source-free amodal panoptic segmentation paper with a useful new benchmark, but the headline numbers are compromised by test-set hyperparameter selection.","tokens_in":28344,"tokens_out":2064,"would_cite":true,"duration_ms":25500,"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":"Source-free panoramic segmentation now matches full-source methods","keywords":["source-free domain adaptation","panoramic segmentation","amodal instance segmentation","panoptic segmentation","pseudo-labeling","class-wise self-tuning thresholds","synthetic-to-real adaptation","occlusion-aware reasoning"],"falsifier":"Take a target panorama set with deliberately degraded or scrambled semantic predictions (e.g., corrupt a fraction of the semantic Thing masks) and rerun UNLOCK while keeping the instance branches untouched: if mAPQ and mAAP drop to or below the source-only baseline as corruption increases, the semantic-revision premise is doing the claimed work. Alternatively, compare UNLOCK against a variant that revises instance masks using a different source of object evidence, such as instance branch self-consistency, on a heavy-distortion target; if that variant matches or beats UNLOCK, the semantic-trust assumption is not necessary.","tokens_in":27287,"feed_emoji":"🌐","tokens_out":4390,"duration_ms":48242,"temperature":0.7,"pith_summary":"Panoramic scene understanding normally needs either labeled panoramic data or access to the labeled pinhole data used to train the source model. This paper introduces a stricter task, source-free occlusion-aware seamless segmentation (SFOASS), in which neither the source images nor the target labels are available during adaptation, and proposes the first solution, UNLOCK. The paper claims that a pinhole-trained model can be adapted to unlabeled 360-degree imagery purely through pseudo-labeling and a mixing strategy, and backs this with state-of-the-art scores of 10.9 in mAAP and 11.6 in mAP, plus a +4.3 gain in mAPQ over the source-only baseline on the real-to-real benchmark. The contribution matters because privacy, storage, and commercial restrictions often make the source dataset inaccessible, and the results suggest that giving up source access need not cost performance.","feed_headline":"Source-free panoramic segmentation now matches full-source methods","feed_subtitle":"UNLOCK adapts pinhole-trained models to 360-degree imagery using pseudo-labels and amodal-driven mixing.","key_machinery":"Two interacting modules carry the argument. Omni Pseudo-Labeling Learning (OPLL) combines class-wise self-tuning thresholds with semantic-to-instance mask revision: the semantic branch's Thing mask is intersected with each instance-level prediction, so objects the local instance branch mislabels can be corrected, and an uncertainty-guided binary cross-entropy loss masks out low-confidence regions so wrong pseudo-labels are not penalized as false background. Amodal-Driven Contextual Learning (ADCL) builds a pool of high-confidence amodal object samples and mixes them into training images, zeroing only the overlapping (potentially occluded) regions while keeping the full amodal mask as the target label, which preserves the complete shape of objects without contaminating scene context.","core_discovery":"The central claim is that source-free adaptation is sufficient for occlusion-aware seamless segmentation on panoramic images. UNLOCK generates omni pseudo-labels from all three branches of the source model, using class-wise self-tuning thresholds that adapt per category rather than one global cutoff, and refines instance-level masks with the semantic branch's Thing mask. It then builds an amodal-driven object pool of the most confident amodal instance predictions and uses a spatial-aware mixing strategy that pastes full object shapes while zeroing only their overlapping regions, so the model sees complete occluded objects without corrupting contextual cues. On KITTI360-APS->BlendPASS the method reaches 26.4 mAPQ, 10.9 mAAP, and 11.6 mAP, surpassing UDA methods that use all 12K source image-label pairs, and on the new AmodalSynthDrive->BlendPASS synthetic-to-real benchmark it also leads across the five segmentation metrics.","pith_inferences":["Beyond the paper: the principle of zeroing only overlapping regions while keeping full amodal masks suggests a general recipe for amodal data augmentation, and applying it to other amodal datasets such as indoor scenes is a direct test.","Beyond the paper: if source-free adaptation matches UDA here, the value of source data may mostly lie in pre-training, and future work could investigate how much source data is actually needed before it stops helping.","Beyond the paper: the reported failure case of vehicles behind sparse fences suggests that structured occluders that break spatial continuity, not just dense ones, are the remaining hard mode; a stress-test suite built from such occluders would be a useful benchmark.","Beyond the paper: the SFOASS task formulation could become a template for other source-free multi-branch dense prediction settings, such as 3D detection with instance and amodal branches."],"forward_implications":["Source-free adaptation can serve as a drop-in replacement when source data is restricted, reducing storage and privacy burdens.","Synthetic-to-real transfer works for occlusion-aware segmentation, so simulators can be used to train panoramic models without real pinhole labels.","The method generalizes beyond OASS: applied to semantic-only panoramic adaptation (C2D benchmark) it improves mIoU to 54.55, indicating the pseudo-labeling strategy transfers.","The class-wise thresholding component addresses the imbalance in source-model confidence across categories, a common failure mode of fixed-threshold self-training.","UNLOCK is backbone-agnostic, showing consistent gains across different architectures, so the adaptation strategy can be composed with stronger base models."],"supporting_citations":[{"why":"Defines the OASS task with its three-branch architecture, the BlendPASS target dataset, the K2B benchmark, and the AoMix mixing baseline that UNLOCK extends.","marker":"[7]"},{"why":"Provides 360SFUDA++, the source-free panoramic segmentation method that UNLOCK extends to instance-level branches and outperforms on both benchmarks.","marker":"[96]"},{"why":"Supplies the KITTI360-APS pinhole source dataset with amodal annotations for the real-to-real adaptation scenario.","marker":"[49]"},{"why":"Supplies the AmodalSynthDrive synthetic pinhole source dataset that enables the newly introduced synthetic-to-real A2B benchmark.","marker":"[58]"},{"why":"Trans4PASS is a distortion-aware panoramic UDA baseline that UNLOCK must beat in the real-to-real and synthetic-to-real comparisons.","marker":"[86]"},{"why":"EDAPS is a domain-adaptive panoptic segmentation baseline used for comparison and reproduced on the new A2B benchmark.","marker":"[56]"},{"why":"DATR is a distortion-aware panoramic UDA baseline used as a comparison method in all five segmentation metrics.","marker":"[94]"},{"why":"Provides the source-free semantic segmentation baseline SFDA that grounds the C2D generalization experiment for UNLOCK.","marker":"[41]"}],"fun_headline_variants":["Source-free UNLOCK beats full-source panoramic segmentation","No source data? UNLOCK still wins at 360° segmentation","UNLOCK masters occlusion-aware 360° segmentation source-free","Adapting pinhole to panorama: no source needed with UNLOCK","UNLOCK: seamless panoramic segmentation without any source data"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that on the target panoramic domain the source model's semantic branch is more trustworthy than its instance-level branches, so instance pseudo-labels can safely be revised by intersecting them with the semantic Thing mask; if a semantic prediction is wrong on heavily distorted or unfamiliar classes, correct instance detections can be discarded or corrupted.","fun_headline_variants_meta":{"raw":{"variants":["Source-free UNLOCK beats full-source panoramic segmentation","No source data? UNLOCK still wins at 360° segmentation","UNLOCK masters occlusion-aware 360° segmentation source-free","Adapting pinhole to panorama: no source needed with UNLOCK","UNLOCK: seamless panoramic segmentation without any source data"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000378,"raw_usage":{"total_tokens":2026,"prompt_tokens":974,"completion_tokens":1052,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":590,"completion_tokens_details":{"reasoning_tokens":969}},"tokens_in":590,"tokens_out":1052,"duration_ms":11068,"temperature":1.0,"reasoning_tokens":969,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T22:31:05.409800+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a target panorama set with deliberately degraded or scrambled semantic predictions (e.g., corrupt a fraction of the semantic Thing masks) and rerun UNLOCK while keeping the instance branches untouched: if mAPQ and mAAP drop to or below the source-only baseline as corruption increases, the semantic-revision premise is doing the claimed work. Alternatively, compare UNLOCK against a variant that revises instance masks using a different source of object evidence, such as instance branch self-consistency, on a heavy-distortion target; if that variant matches or beats UNLOCK, the semantic-trust assumption is not necessary.","supporting_citations":[{"cited_title":"Vasilakos, and Lin Wang","cited_arxiv_id":null,"evidence_quote":"Provides 360SFUDA++, the source-free panoramic segmentation method that UNLOCK extends to instance-level branches and outperforms on both benchmarks."},{"cited_title":"Amodal panoptic seg- mentation","cited_arxiv_id":null,"evidence_quote":"Supplies the KITTI360-APS pinhole source dataset with amodal annotations for the real-to-real adaptation scenario."},{"cited_title":"AmodalSynthDrive: A syn- thetic amodal perception dataset for autonomous driving","cited_arxiv_id":null,"evidence_quote":"Supplies the AmodalSynthDrive synthetic pinhole source dataset that enables the newly introduced synthetic-to-real A2B benchmark."},{"cited_title":"Bending reality: Distortion-aware transformers for adapting to panoramic semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Trans4PASS is a distortion-aware panoramic UDA baseline that UNLOCK must beat in the real-to-real and synthetic-to-real comparisons."},{"cited_title":"EDAPS: Enhanced domain-adaptive panoptic segmentation","cited_arxiv_id":null,"evidence_quote":"EDAPS is a domain-adaptive panoptic segmentation baseline used for comparison and reproduced on the new A2B benchmark."},{"cited_title":"Look at the neighbor: Distortion-aware unsupervised domain adaptation for panoramic semantic segmentation","cited_arxiv_id":null,"evidence_quote":"DATR is a distortion-aware panoramic UDA baseline used as a comparison method in all five segmentation metrics."},{"cited_title":"Source-free domain adaptation for semantic segmentation","cited_arxiv_id":null,"evidence_quote":"Provides the source-free semantic segmentation baseline SFDA that grounds the C2D generalization experiment for UNLOCK."}],"review_version":1}