{"id":"08a8b576-47c4-46c1-8a21-8da9bbf763c3","arxiv_id":"2506.14686","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FocalClick-XL, a three-subnet extension of FocalClick, achieves state-of-the-art click-based interactive segmentation and supports boxes, scribbles, and coarse masks through a single prompting layer.","lead":"This paper rebuilds the interactive segmentation model FocalClick as a three-stage pipeline that separately handles scene context, object location, and fine detail, and it shows the system works with clicks, boxes, scribbles, and rough masks. It reports fewer user clicks than prior systems on standard benchmarks and adds new test sets and protocols for evaluating non-click interactions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table 7 itself contradicts the 'state-of-the-art' claim: SimpleClick-H has lower NoC85 on SBD than FocalClick-XL-H.","rationale":"The reader's verdict is CONDITIONAL, and my concern reinforces rather than overturns that verdict, so I recommend UNCHANGED. I am not claiming the method is wrong; rather, the strongest claim as stated is not supported by the paper's own evidence. The reader's weakest_assumption concerned the fidelity of simulated scribbles, boxes, and coarse masks to real human input; that is a real concern for the 'unified' and 'adaptability' contributions. However, the most load-bearing assertion identified in the verdict is about click-based state-of-the-art, and the most direct check of that assertion is internal to Table 7: SimpleClick-H already beats FocalClick-XL-H on SBD NoC85. This is a concrete, non-speculative inconsistency that does not require external assumptions about user behavior. The missing SAM 2 comparison noted by the reader is also relevant, but the SBD result is stronger because it comes from the paper's own table. My proposed check is a straightforward re-evaluation with the identical protocol, plus an optional SAM 2 run to test the broader SOTA claim. If the SBD numbers reproduce, the paper should either soften the headline or report SOTA by a defined aggregate metric. This is a claim-revision issue, not a rejection of the method, and it leaves the reader's CONDITIONAL verdict unchanged.","tokens_in":23111,"tokens_out":17241,"duration_ms":181442,"concrete_test":"Re-run the Table 7 evaluation of SimpleClick-H and FocalClick-XL-H on SBD with the exact same simulator and code path (same click placement, same upper bound, same SBD test split) to confirm the NoC85 difference. If the numbers reproduce (2.85 vs 3.32), revise the abstract and Sec. 7.1 to say FocalClick-XL achieves state-of-the-art on GrabCut, Berkeley, and DAVIS, and competitive on SBD, or state SOTA by average NoC. Additionally, report the same comparison for SAM 2 under this protocol to test the unqualified SOTA claim against current baselines.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central assertion is that FocalClick-XL achieves state-of-the-art click-based segmentation and, per the reader's strongest claim, fewer clicks than prior methods on GrabCut, Berkeley, SBD, and DAVIS. The paper's own Table 7 does not support this on SBD: SimpleClick-H, trained only on COCO+LVIS, reports SBD NoC85 = 2.85 and NoC90 = 4.70, whereas FocalClick-XL-H, trained on the larger Combined Dataset, reports SBD NoC85 = 3.32 and NoC90 = 5.51. Thus a prior method outperforms FocalClick-XL on one of the four headline benchmarks. The claim may still hold on average or on the other three datasets, but the unqualified 'state-of-the-art on click-based benchmarks' is internally contradicted. This matters because the abstract and Sec. 7.1 use this claim to motivate the entire pipeline; if the comparison were recomputed with identical protocol, the headline would need to be narrowed.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript extends the FocalClick interactive-segmentation framework into FocalClick-XL, a three-subnet pipeline that decomposes the task into context-level, object-level, and detail-level processing. The Context-Net is built on a SAM encoder with adapters, the Object-Net is a SegFormer-style model that fuses interaction prompts, and the Detail-Net refines low-confidence regions and can predict alpha mattes. The paper reports state-of-the-art click-based results on GrabCut, Berkeley, SBD, and DAVIS, and shows competitive performance for boxes, scribbles, and coarse masks under newly proposed or newly evaluated protocols, including a new DAVIS-585 benchmark for interactive mask correction and a deterministic scribble simulator.","tokens_in":23359,"tokens_out":6518,"duration_ms":63133,"significance":"If the reported results hold, FocalClick-XL is a substantial contribution to interactive segmentation: it combines large-scale pretraining with a coarse-to-fine design, supports multiple interaction formats in one architecture, and introduces reusable evaluation protocols for non-click interactions. The paper's strengths include comparisons against standard external benchmarks, stepwise ablations showing the contribution of each subnet, and the explicit goal of sharing context- and detail-level knowledge across interaction types. The main caveats are that the headline 'state-of-the-art' claim is contradicted by the paper's own SBD numbers, statistical reliability is not established, and the training setup for the XL model is underreported. These issues are fixable and do not invalidate the core methodology.","major_comments":[{"comment":"The unqualified 'state-of-the-art on click-based benchmarks' claim in the abstract and §7.1 is contradicted by the paper's own Table 7. SimpleClick-H trained on COCO+LVIS achieves SBD NoC85 = 2.85 and NoC90 = 4.70, while FocalClick-XL-H trained on the larger Combined Dataset achieves SBD NoC85 = 3.32 and NoC90 = 5.51. FocalClick-XL is better on GrabCut, Berkeley, and DAVIS, but the SOTA claim should be qualified (e.g., 'on most click-based benchmarks' or 'excluding SBD') and the discrepancy should be discussed rather than glossed over.","section":"§7.1, Table 7 (SBD block)"},{"comment":"All quantitative tables report point estimates without error bars, confidence intervals, or significance tests. GrabCut and Berkeley contain only 50 and 100 images, yet several headline comparisons rest on differences of 0.08–0.16 clicks (e.g., FocalClick-XL-B vs FocalClick-XL-H on GrabCut NoC85: 1.30 vs 1.22). Please report per-image variance or paired statistical tests for the central claims, especially in Table 7.","section":"§7.1, Tables 7–8"},{"comment":"The training setup for FocalClick-XL is not reported. Whereas §6.1 gives dataset, epoch count, batch size, optimizer, learning-rate schedule, and augmentation for basic FocalClick, §7 does not state the Combined Dataset composition, number of iterations, batch size, learning-rate schedule, input resolutions for the subnets, or the order of click-pretraining versus transfer tuning. Without these details the headline results cannot be reproduced.","section":"§7"},{"comment":"The claim that the decomposition 'allows each subnet to undergo scaled pretraining with independent data and supervision' is not directly tested. Table 9 ablates the presence of the three subnets but keeps the pretraining and data scheme fixed; there is no experiment comparing independent per-subnet pretraining with a jointly trained end-to-end counterpart. Please provide such an ablation, or temper the 'maximizing its effectiveness' claim.","section":"§4.1 and Table 9"},{"comment":"The scribble, box, and coarse-mask evaluations rest entirely on simulated user interactions, and the only human evidence is a one-sentence user study in §3.2 that is not described in terms of protocol or sample size. Since the paper proposes DAVIS-585 and the scribble protocol as new benchmarks, the simulators should be validated against a small set of real human annotations, or the results should be explicitly framed as simulator-based.","section":"§5.1–§5.2"}],"minor_comments":[{"comment":"The text says 'Adam optimizer of β1 = 0.9, β1 = 0.999'; the second should presumably be β2 = 0.999.","section":"§6.1"},{"comment":"The header lists 'NoC 90' for Berkeley but not 'NoC 85'; please check the column alignment and header formatting so each column is unambiguous.","section":"Table 7"},{"comment":"The text says 'We combine these four strategies' but only three meta-simulators are described (Bezier, axial, boundary scribbles); clarify what the fourth strategy is.","section":"§5.1"},{"comment":"The text says adapters are added 'after each transformer layer,' while Table 14 uses the label 'Each stage'; specify whether these refer to the same placement and report the number of adapters.","section":"§4.1 vs Table 14"},{"comment":"The text promises NoF2085/90 for scribble evaluation, but Table 10 reports only NoS; either add the failure numbers or remove the promise.","section":"§5.2"},{"comment":"There are several typos: 'Comaprisons' in the Table 7 caption, 'Pertub-Level' in Table 12, 'Promtping Layer' in §5.1, and 'archives impressive performance' in §7.1.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":"The paper does not state whether code and the new benchmarks will be released; for reproducibility of DAVIS-585 and the scribble protocol, the editor may want to ask the authors to commit to a release. The SBD discrepancy in Table 7 should be resolved before the paper can be accepted, as the abstract currently overclaims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is the real deal as an engineering contribution, but the paper's headline claim is internally contradicted. Table 7 shows SimpleClick-H gets SBD NoC85 2.85 / NoC90 4.70 with COCO+LVIS training, while FocalClick-XL-H gets 3.32 / 5.51 on the larger Combined Dataset. So \"state-of-the-art on click-based benchmarks\" is not true across the four headline datasets; it's true on GrabCut, Berkeley, and DAVIS, but not SBD. That's a fixable claim-narrowing problem, not a fatal flaw.\n\nWhat's actually new: the three-subnet decomposition (Context-Net, Object-Net, Detail-Net) with independent large-scale pretraining per subnet, the unified 2-channel prompting layer that lets one model take clicks, boxes, scribbles, and coarse masks, and two new evaluation artifacts: DAVIS-585 for mask correction and the deterministic scribble simulator. The ablations are honest, and each component shows a steady gain. Progressive magnification keeps online speed competitive; the box and scribble numbers look solid, and the coarse-mask refinement results are impressive even at low initial IoUs. This is careful, reproducible-looking work in the sense that the method is described precisely enough to reimplement, and the benchmarks are well-defined.\n\nSoft spots, in order of real weight. First, the SBD contradiction above; the abstract and Sec. 7.1 should be narrowed or the authors need to explain why SBD behaves differently. Second, SAM 2 is absent from the comparisons; a 2025 paper claiming click SOTA needs to include it. Third, no code or models are released, so independent verification is impossible; the reader's conditional verdict rides on that. Fourth, no error bars or significance tests anywhere, which matters on small sets like GrabCut (50 images). Fifth, the alpha-matte claim is not quantitatively evaluated — no matting metrics, no comparisons, just qualitative images. That claim should be either supported or cut. Finally, the FocalClick-XL training details in Sec. 7 are much thinner than the basic FocalClick setup; exact data mix, schedule, and hyperparameters need to be specified.\n\nThe simulated interaction protocols (scribbles, defective masks) are author-defined and could differ from real annotators, but that's a minor concern because they are used as internal test grounds, not as fitted targets; still, a human-user study would strengthen the DAVIS-585 and scribble claims.\n\nBottom line: this paper deserves a serious referee. It's a solid within-subfield contribution with useful artifacts and a clear architecture. I'd recommend major revision: fix the claim, add SAM 2, release code, quantify the matting, add error bars. If those land, it's a good paper. For a reading group, I'd bring it; the benchmark proposals alone are worth discussing.","headline":"Strong engineering and useful new benchmarks, but the paper's own Table 7 contradicts the 'state-of-the-art on click-based benchmarks' claim on SBD.","tokens_in":23880,"tokens_out":3542,"would_cite":true,"duration_ms":31405,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"FocalClick-XL splits interactive segmentation into three pretrained subnets and claims fewer clicks than SAM on every standard benchmark.","keywords":["interactive segmentation","click-based segmentation","scribble interaction","box prompting","coarse mask refinement","alpha matting","coarse-to-fine refinement","progressive magnification"],"falsifier":"Run a human annotation study in which real users correct masks on the same images with free-form scribbles and clicks, comparing their interaction counts against the simulator-driven NoS/NoC numbers; a systematic gap beyond the reported noise would refute the transfer claim, as would a failure of the deterministic scribble simulator to predict human placement.","tokens_in":22906,"feed_emoji":"🖱️","tokens_out":5279,"duration_ms":49045,"temperature":0.7,"pith_summary":"FocalClick-XL aims to show that interactive segmentation is better built as three separately trained subnets — one for scene context, one for the target object, and one for fine detail — than as a single monolithic prompt-to-mask model. Each subnet is pretrained on data suited to its job, and only the object-level subnet needs to know what kind of interaction the user gave. If the paper is right, one model can serve clicks, boxes, scribbles, and coarse masks while matching or beating specialized methods on click benchmarks and even predicting alpha mattes for hair and fur. This matters because real annotation tools must handle varied input formats and fine boundaries at low latency, and the paper supplies a concrete architecture plus benchmarks for measuring both.","feed_headline":"Three subnets beat SAM at click segmentation","feed_subtitle":"Context, object, and detail subnets pretrain separately, then unify clicks, boxes, scribbles, coarse masks, and mattes.","key_machinery":"The load-bearing mechanism is the three-subnet decomposition. Context-Net reuses a SAM encoder with per-block adapters and runs once per image; Object-Net, a SegFormer-style network, fuses the interaction prompt, encoded as a two-channel positive/negative map, with the zoomed target patch and RoI-aligned context features; Detail-Net, a MobileNetV2 U-shape, takes a tri-map of confident foreground, confident background, and uncertain regions and predicts a refined mask or alpha matte whose per-pixel values encode transparency. Progressive magnification crops small patches around the object and then around low-confidence regions, which keeps the cascaded pipeline fast. Progressive Merge preserves user-annotated details outside the region the new interaction is meant to change.","core_discovery":"The paper claims that decomposing interactive segmentation into context, object, and detail levels, with each subnet pretrained independently, yields state-of-the-art click-based performance and broad interaction support. Concretely, FocalClick-XL reports lower average numbers of clicks to reach 85% and 90% IoU than prior methods, including SAM and SAM-HQ, on GrabCut, Berkeley, SBD, and DAVIS. The same model transfers to box, scribble, and coarse-mask inputs by retraining only a small prompting layer, and the detail subnet trained on matting data outputs alpha values, so prediction is not limited to binary masks. The authors also introduce evaluation protocols and new benchmarks, including DAVIS-585 for correcting flawed initial masks and a deterministic scribble simulator.","pith_inferences":["If real annotators' scribbles and coarse masks resemble the simulated ones, the prompting-layer-only transfer suggests the same decomposition could absorb language or reference-image prompts with a similarly small per-interaction cost; the paper does not test this.","The claim that only the object-level subnet is interaction-sensitive implies that interaction types share context and detail knowledge; this could be tested by training on a mixed-interaction dataset and checking whether Object-Net capacity becomes a bottleneck.","The DAVIS-585 correction benchmark could be adopted widely, since real annotation workflows often start from prior predictions, though the paper itself only simulates defects on ground-truth masks."],"forward_implications":["A new interaction format can be supported by training only its prompting layer, leaving Context-Net and Detail-Net frozen.","Because the detail subnet is trained on matting data, the same pipeline outputs alpha mattes for transparent structures, not just binary masks.","The DAVIS-585 benchmark turns mask correction into a measurable task, showing large click savings when starting from a flawed initial mask.","The deterministic scribble simulator replaces random scribble generation with a reproducible protocol, enabling automatic NoS evaluation."],"supporting_citations":[{"why":"Supplies the large-scale pretrained encoder that Context-Net adapts, and serves as the main baseline with its click and box results.","marker":"[2]"},{"why":"Provides the iterative training scheme, normalized focal loss, mask-guidance design, and click-simulation protocol that the pipeline builds on.","marker":"[7]"},{"why":"The conference-version coarse-to-fine pipeline (Target Crop, Focus Crop, Progressive Merge) that this work scales up and extends.","marker":"[9]"},{"why":"One of the four click-based benchmarks on which the state-of-the-art claim is measured.","marker":"[15]"},{"why":"One of the four click benchmarks and the source of images and masks for the new DAVIS-585 correction benchmark.","marker":"[54]"},{"why":"Used both as a click-based benchmark and as part of the training and evaluation data for segmentation.","marker":"[55]"},{"why":"Object-centric training data used to pretrain Object-Net together with the Context-Net adapters.","marker":"[60]"},{"why":"Object-centric training data combined with COCO for object-level pretraining.","marker":"[61]"},{"why":"SAM variant with interaction-agnostic refinement that is a baseline for clicks, boxes, and high-quality segmentation.","marker":"[34]"}],"fun_headline_variants":["Multi-subnet model beats SAM on click segmentation","FocalClick-XL: subnets for context, object, detail unify interactions","Decomposed subnets beat SAM at clicks, handle boxes and scribbles","Three subnets pretrain separately, unify clicks, boxes, and mattes","Interactive segmentation via context, object, and detail subnets"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's central results rest on the assumption that the simulated users — max-error clicks, Bezier/axial/boundary scribbles, and eroded or dilated coarse masks — behave like real annotators, so performance under simulation transfers to practice.","fun_headline_variants_meta":{"raw":{"variants":["Multi-subnet model beats SAM on click segmentation","FocalClick-XL: subnets for context, object, detail unify interactions","Decomposed subnets beat SAM at clicks, handle boxes and scribbles","Three subnets pretrain separately, unify clicks, boxes, and mattes","Interactive segmentation via context, object, and detail subnets"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000607,"raw_usage":{"total_tokens":2817,"prompt_tokens":925,"completion_tokens":1892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":1798}},"tokens_in":541,"tokens_out":1892,"duration_ms":14197,"temperature":1.0,"reasoning_tokens":1798,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T19:47:59.949964+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a human annotation study in which real users correct masks on the same images with free-form scribbles and clicks, comparing their interaction counts against the simulator-driven NoS/NoC numbers; a systematic gap beyond the reported noise would refute the transfer claim, as would a failure of the deterministic scribble simulator to predict human placement.","supporting_citations":[{"cited_title":"Focalclick: Towards practical interactive image segmentation","cited_arxiv_id":null,"evidence_quote":"The conference-version coarse-to-fine pipeline (Target Crop, Focus Crop, Progressive Merge) that this work scales up and extends."},{"cited_title":"” grabcut” interactive foreground extraction using iterated graph cuts.TOG, 2004","cited_arxiv_id":null,"evidence_quote":"One of the four click-based benchmarks on which the state-of-the-art claim is measured."},{"cited_title":"A benchmark dataset and evaluation methodology for video object segmentation","cited_arxiv_id":null,"evidence_quote":"One of the four click benchmarks and the source of images and masks for the new DAVIS-585 correction benchmark."},{"cited_title":"Semantic contours from inverse detectors","cited_arxiv_id":null,"evidence_quote":"Used both as a click-based benchmark and as part of the training and evaluation data for segmentation."},{"cited_title":"Microsoft coco: Common objects in context","cited_arxiv_id":null,"evidence_quote":"Object-centric training data used to pretrain Object-Net together with the Context-Net adapters."},{"cited_title":"Lvis: A dataset for large vocabulary instance segmentation","cited_arxiv_id":null,"evidence_quote":"Object-centric training data combined with COCO for object-level pretraining."}],"review_version":2}