{"id":"25a07982-5d01-43ef-b92b-21fff82d6e6f","arxiv_id":"2505.13099","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Formula-generated hollow masks (InsCore) outperform real-image pre-training datasets and fine-tuned SAM on five industrial instance segmentation benchmarks using only 100,000 synthetic images.","lead":"The paper introduces InsCore, a synthetic pre-training dataset of 100,000 formula-generated images with hollow masks, and reports that it beats COCO, ImageNet-21k, and fine-tuned SAM on five industrial instance segmentation benchmarks. A generalist should read it because it suggests a license-free, annotation-free path to industrial segmentation models, though the headline margins are smaller than the abstract implies.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"InsCore's recipe is tuned on the same five benchmarks used for evaluation, so the reported gains may not transfer to new industrial domains; a held-out test is needed.","rationale":"The reader's weakest assumption is exactly the hyperparameter/generalization issue, and I agree it is the most load-bearing. The controlled comparison to COCO (same Mask R-CNN Swin-B architecture) is the cleanest evidence for the core claim, but the margin is only 0.4 mAP on average; the larger 6.2-point margin over SAM comes from a non-architecture-matched comparison using mIoU. Those weaknesses are secondary to the fact that the method's free parameters were chosen by maximizing performance on the same five test sets. A foundation-model claim requires at least one held-out domain, and leave-one-dataset-out selection is a concrete, feasible check that would distinguish a general recipe from benchmark-tuned settings. I therefore keep the reader's CONDITIONAL verdict unchanged.","tokens_in":13124,"tokens_out":7146,"duration_ms":77591,"concrete_test":"Run leave-one-dataset-out validation over the five Table 1 datasets: for each held-out dataset, select the InsCore hyperparameters (occlusion rate, mask offset, dataset size) using the Section 4.3 ablations on the remaining four datasets only, then fine-tune Mask R-CNN Swin-B with the Table 2 protocol and evaluate on the held-out dataset. Compare the average held-out mAP against the identical procedure with COCO pre-training. If InsCore does not beat COCO on held-out domains, the reported advantage is benchmark-specific rather than domain-agnostic.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The weakest point is that the InsCore recipe is selected and evaluated on the same five datasets. Section 4.3 ablates occlusion rate (Table 5), mask precision (Table 6), and dataset size (Table 7), and picks occlusion=100, mask offset=0, and 100k images; these are exactly the settings used in the headline comparisons of Tables 2 and 3, with test performance measured on the same benchmarks. Because no held-out industrial dataset or nested validation is reported, the average gains (46.0 vs 45.6 mAP over COCO; 75.0 vs 68.8 mIoU over fine-tuned SAM) may partly reflect choosing hyperparameters that maximize scores on these five test sets. The paper's claim to be a 'domain-agnostic, general-purpose' foundation model (Section 4.1, abstract) requires transfer beyond the five chosen domains; the current experiments cannot support that. If the fixed recipe fails on a new industrial domain, the central claim that formula-generated hollow masks replace real-image pre-training for industry generally is not established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper introduces InsCore, a fully synthetic pre-training dataset for industrial instance segmentation. InsCore images are generated by a formula-driven process: each image contains randomly placed hollow masks defined as the region between nested RCDB-style contour polygons, with back-to-front rendering producing complex occlusion hierarchies and 256 pseudo-class labels. No real images or manual annotations are used. The authors pre-train Mask R-CNN with a Swin-B backbone on 100k InsCore images and fine-tune on five industrial datasets (Endoscapes, LIVECell, SpaceNet2, Industrial-iSeg, LogiSeg). They report that InsCore pre-training outperforms COCO and ImageNet-21k pre-training by 0.4 mAP on average (Table 2), outperforms fine-tuned SAM by 6.2 mIoU points (Table 3), and improves COCO fine-tuning over ImageNet-21k (Table 4) and RTMDet detection over ImageNet-1k (Table 8). The paper claims these results position InsCore as a license-free, domain-agnostic foundation model for industrial segmentation.","tokens_in":13296,"tokens_out":8780,"duration_ms":78439,"significance":"If the results hold, the contribution is practically significant: InsCore is a scalable, license-free alternative to real-image pre-training for instance segmentation, with strong data efficiency (0.1M images vs 11M in SA-1B) and positive transfer across five industrial domains and on COCO. The method's reliance on a single generation formula and its delivery of precise pixel annotations with no human labor are genuine strengths. The experimental design covers a broad set of industrial domains, and the RTMDet experiments in the appendix extend the claim beyond Mask R-CNN. However, the current manuscript contains internal inconsistencies and lacks the controlled, statistically grounded evaluation needed to substantiate the central claims; these issues must be resolved before the paper's conclusions can be accepted.","major_comments":[{"comment":"The reported Endoscapes (ES) score for InsCore is inconsistent across tables: Table 2 gives 29.7 mAP, while the same setting (occlusion rate 100, pixel offset 0, 100k images) in Tables 5, 6, and 7 gives 37.1 mAP. If the 37.1 value were used, the InsCore average in Table 2 would be approximately 47.5 rather than 46.0, changing the comparison with COCO from +0.4 to +1.9 mAP points. This discrepancy directly affects the headline claim and must be resolved before the results can be interpreted.","section":"Tables 2, 5-7 (Sections 4.2-4.3)"},{"comment":"The generation hyperparameters (occlusion rate 100, mask offset 0, dataset size 100k) are selected by ablating on the same five industrial datasets that are later used to report the main comparisons in Tables 2 and 3. Since no held-out industrial dataset or nested validation is used, the reported averages are optimistically biased by selection on the evaluation data. The claim that InsCore is a 'domain-agnostic, general-purpose' pre-training method (Section 4.1 and the abstract) requires validation on at least one dataset not used in the ablations, or a nested validation protocol.","section":"Section 4.3, Tables 5-7"},{"comment":"The comparison with SAM is confounded by architecture and prompting protocol: InsCore is evaluated with Mask R-CNN (Swin-B backbone) without prompts, whereas SAM is a ViT-B model that receives ground-truth bounding box prompts. The mIoU comparison therefore does not isolate the contribution of pre-training data. Additionally, zero-shot SAM outperforms fine-tuned SAM on three of the four datasets, which suggests the fine-tuned baseline is not a well-calibrated reference for this claim. A controlled comparison (for example, the same detector head with a SAM-initialized backbone, or a prompt-free SAM variant) is needed to support the claim that InsCore pre-training outperforms fine-tuned SAM.","section":"Section 4.2, Table 3"},{"comment":"All experiments report point estimates from a single run, with no error bars, multiple seeds, or statistical tests. Given that the average mAP advantage over COCO is only 0.4 points and that InsCore is lower than COCO on Endoscapes (29.7 vs 32.0) and LogiSeg (95.1 vs 95.4) in Table 2, the claimed superiority over COCO is not established with demonstrated significance. The authors should report variance across at least three seeds and provide confidence intervals or a paired test for the key comparisons.","section":"Tables 2-7"},{"comment":"The abstract's 'average improvement of 6.2 points in instance segmentation performance' conflates two different metrics and baselines: the 6.2-point gain is mIoU over fine-tuned SAM (Table 3), while the gain over COCO in mAP is 0.4 points (Table 2). The '14.5% accuracy improvement' stated in Figure 1 also lacks a clear definition; 75.0 vs 68.8 mIoU is a relative improvement of about 9.0%, not 14.5%. The claims should be restated with the metric and baseline for each number.","section":"Abstract and Figure 1"}],"minor_comments":[{"comment":"The exact value ranges or distributions for the shape parameters r_k, lw_k, o_k, and lambda_k are not specified in the paper. Since the method is a data generation recipe, these details are necessary for reproducibility; the authors should report the exact sampling distributions used in the experiments.","section":"Section 3.2, Eqs. (1)-(4)"},{"comment":"The table marks COCO as permitting commercial use, but COCO is composed of images from Flickr under varying licenses and is typically distributed for research purposes. The claim should be qualified or the table footnoted to clarify the license status.","section":"Table 2, 'Com. Use' column"},{"comment":"The qualitative claim that InsCore shows 'the steepest and most stable convergence' is not quantified; for LIVECell the curves appear qualitatively similar across pre-training methods. A quantitative convergence metric (for example, loss at a fixed iteration or area under the curve) would strengthen the claim.","section":"Appendix A, Figure 3"},{"comment":"Table 3 averages over four datasets because SAM was not fine-tuned on LIVECell, while Table 2 averages over five datasets. The text should explicitly state that the mIoU averages are computed over different dataset subsets to avoid misleading comparisons.","section":"Section 4.2, Table 3"}],"recommendation":"major_revision","confidential_remarks":"The internal inconsistency in the Endoscapes scores between Table 2 and Tables 5-7 is the most serious issue and should be verified first. If the correct value is 37.1, the average comparisons in the abstract and conclusions change materially; if the correct value is 29.7, the ablation tables need substantial correction. The paper's central idea is interesting and the synthetic-generation mechanism is well-motivated, but the current evaluation does not yet support the breadth of the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is not a rehash: InsCore is the first FDSL dataset aimed at instance segmentation, and the hollow-mask-with-occlusion design is new relative to SegRCDB. Second, the central result is real but smaller than advertised: the 6.2 point improvement is mIoU over fine-tuned SAM, not the mAP comparison with COCO, which is 0.4. Read the tables carefully.\n\nWhat the paper does well: the generation procedure is clearly specified, the experiments cover five genuinely different industrial domains, and the data efficiency story (100k formula-generated images vs 14M ImageNet-21k) is worth taking seriously. The appendix loss curves and the RTMDet sanity check in Appendix B are the kind of evidence that makes the core claim plausible. If I needed a license-free pretraining route for a narrow-domain segmentation task, I would try this recipe.\n\nWhere it is soft. The SAM comparison mixes architecture (Mask R-CNN Swin-B vs SAM ViT-B), training protocol, and metric (mAP vs mIoU), and prompts SAM with ground-truth boxes, which is favorable to SAM; the win is still consistent, but it is not a clean comparison. No error bars are reported anywhere, and some differences (e.g., 46.0 vs 45.6 on average over COCO) are within what random seed variation could produce. The more serious issue is the one the stress-test note hits: the key hyperparameters—occlusion rate, mask precision, dataset size—are ablated on the same five datasets used in the headline tables, and the best settings on those ablations are then used for the main results. That makes the reported gains partly the product of tuning to those benchmarks. The 'domain-agnostic, general-purpose' claim in Section 4.1 needs at least one held-out industrial dataset or a nested validation procedure. No code or data is released, which hurts reproducibility.\n\nWho this is for: people working on synthetic pretraining, FDSL, or industrial segmentation. It deserves a serious referee, not a desk reject. In review, I would ask for clearer metric reporting, variance estimates, a held-out transfer check, and release of the generator.","headline":"A credible extension of FDSL to instance segmentation with a genuinely new hollow-mask recipe, but the headline margin is overstated and the recipe is tuned on its own test benchmarks.","tokens_in":13878,"tokens_out":2969,"would_cite":false,"duration_ms":29438,"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":"A 100,000-image synthetic dataset of hollow masks outperforms COCO, ImageNet-21k, and fine-tuned SAM as pre-training for industrial instance segmentation.","keywords":["instance segmentation","synthetic pre-training","formula-driven supervised learning","hollow masks","industrial datasets","occlusion modeling","domain gap","data efficiency"],"falsifier":"Pre-train Mask R-CNN with a Swin-B backbone on the exact InsCore recipe (100k images, occlusion rate 100, zero mask offset, 512px), freeze it, fine-tune on a sixth industrial dataset from a domain not in the paper's ablation, and compare against COCO and ImageNet-21k initializations; if the InsCore-initialized model does not match or beat those baselines on average, the domain-agnostic claim is refuted and the reported gains are benchmark-specific.","tokens_in":12856,"feed_emoji":"🏭","tokens_out":10394,"duration_ms":91218,"temperature":0.7,"pith_summary":"InsCore pre-training uses 100,000 formula-generated images, each filled with hollow contour masks, and no real photographs or manual annotations. The paper claims this is enough to beat established real-image pre-training on industrial instance segmentation: averaged over five datasets in medical, biomedical, remote sensing, manufacturing, and logistics, InsCore reaches 75.0 mIoU versus 68.8 for fine-tuned SAM, and 46.0 mAP versus 45.6 for COCO. If true, the practical consequence is that a license-free, commercially usable vision foundation model for industry can be built from a single mathematical formula rather than from large scraped image collections. The underlying idea is that what transfers to industrial data is not appearance but the structure of occlusion: dense, hierarchical, precisely labeled masks.","feed_headline":"Synthetic hollow masks beat real-image pretraining for industry","feed_subtitle":"100k license-free synthetic images beat COCO, ImageNet-21k, and fine-tuned SAM on industrial segmentation.","key_machinery":"The object that carries the argument is the hollow instance mask $S_k = R_{\\text{out},k} \\setminus R_{\\text{in},k}$, the region between the outer and inner contours of a formula-generated shape. The shape itself begins as a regular $n_k$-gon and is expanded radially through successive polygons with Perlin-noise perturbations, producing diverse non-rigid contours; instances are sorted back-to-front, and later instances occlude earlier ones, so only visible pixels receive labels. This single construction packages the paper's three industrial priors—complex occlusion, dense hierarchical masks, and shape diversity—into a fully supervised pixel-level task, making the synthetic dataset a drop-in replacement for real-image pre-training.","core_discovery":"The central discovery is that the hollow mask, defined as the ring between the outermost and innermost of a stack of recursively generated nested polygons, is a sufficient pre-training signal for industrial instance segmentation. Each hollow shape is an instance; shapes are pasted back-to-front and an instance's visible region is its ring minus the outer bounds of every later-drawn shape, so the ground-truth labels encode occlusion order exactly. Trained on these masks, Mask R-CNN with a Swin-B backbone outperforms the same architecture initialized on ImageNet-21k, COCO, and several synthetic datasets, and also beats fine-tuned SAM on four of five industrial benchmarks even though SAM had 110 times more pre-training images. On general COCO instance segmentation, InsCore pre-training surpasses both ImageNet-21k and SegRCDB. The authors interpret this as evidence that complex occlusion and precise mask boundaries, not real-image texture or semantics, carry the transferable representation.","pith_inferences":["If the hollow-mask prior is the transferable signal, the same recipe should boost other dense prediction tasks, such as panoptic segmentation or video instance segmentation; the appendix's near-parity with ImageNet-1k on RTMDet detection already points that way.","A cleaner test of domain-agnosticism would freeze the generation hyperparameters and evaluate on an unseen sixth industrial domain, since the current settings were selected by ablating on the same five benchmarks.","The structure-not-texture interpretation can be tested by rendering the hollow masks with random interior textures; if downstream mAP stays flat, appearance truly is irrelevant.","A practical hybrid would be to use InsCore to initialize a promptable segmenter's decoder, coupling the synthetic occlusion prior with boundary refinement; the paper does not explore this."],"forward_implications":["Industrial model builders can pre-train on synthetic data without collecting, curating, or licensing real images, eliminating the ImageNet commercial-use restriction that motivated the work.","Because InsCore beats fine-tuned SAM using 1/110 of SA-1B's data, large web-scraped foundation models are not a prerequisite for strong industrial segmentation.","The same 100k-image pre-training helps on general COCO instance segmentation, beating ImageNet-21k and SegRCDB, so the benefit is not confined to industrial imagery.","Synthetic labels' exactness is a real advantage: offsetting masks by 30 pixels drops 7.9 mAP points on the manufacturing benchmark, showing that fuzzy annotations would erode the gain.","Data volume has diminishing returns: going from 100k to 400k images does not improve average performance, so InsCore's effectiveness is a matter of design, not scale."],"supporting_citations":[{"why":"This supplies the Formula-Driven Supervised Learning contour-generation method whose nested polygons InsCore turns into hollow masks.","marker":"[17]"},{"why":"This is the synthetic semantic-segmentation baseline and parameter source that InsCore extends to instance-level hollow masks.","marker":"[29]"},{"why":"This defines the SAM/SA-1B baseline and its 11M-image scale, the efficiency contrast for the 100k-image claim.","marker":"[18]"},{"why":"This provides both the strongest real-image instance-segmentation pre-training baseline and the general-object evaluation benchmark.","marker":"[20]"},{"why":"This is the standard real-image pre-training baseline and the license-restricted dataset the authors aim to replace.","marker":"[8]"},{"why":"This supplies the medical-domain benchmark with complex tool-tissue occlusion.","marker":"[27]"},{"why":"This supplies the dense hierarchical cell-mask benchmark where InsCore shows its largest gain over COCO.","marker":"[10]"},{"why":"This supplies the remote-sensing benchmark of dense building masks.","marker":"[33]"},{"why":"This supplies the manufacturing-domain benchmark of production-line defects.","marker":"[19]"},{"why":"This supplies the logistics-domain benchmark of warehouse and truck-yard objects.","marker":"[23]"}],"fun_headline_variants":["100k synthetic hollow masks outdo 11M SAM on industry","License-free synthetic pretraining beats COCO and SAM","Hollow masks: 100k synthetic images, 6.2 pt industrial boost","Synthetic-only pretraining: better than SAM, 110x fewer images","InsCore: hollow masks beat real-image pretraining for industry"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the specific InsCore recipe—hollow masks at 100k images, occlusion rate 100, and zero mask offset—transfers to industrial data beyond the five benchmarks on which those settings were tuned; if it does not, the claim that InsCore is a domain-agnostic foundation model collapses.","fun_headline_variants_meta":{"raw":{"variants":["100k synthetic hollow masks outdo 11M SAM on industry","License-free synthetic pretraining beats COCO and SAM","Hollow masks: 100k synthetic images, 6.2 pt industrial boost","Synthetic-only pretraining: better than SAM, 110x fewer images","InsCore: hollow masks beat real-image pretraining for industry"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000552,"raw_usage":{"total_tokens":2670,"prompt_tokens":1019,"completion_tokens":1651,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":1558}},"tokens_in":635,"tokens_out":1651,"duration_ms":10018,"temperature":1.0,"reasoning_tokens":1558,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:20:07.305509+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Pre-train Mask R-CNN with a Swin-B backbone on the exact InsCore recipe (100k images, occlusion rate 100, zero mask offset, 512px), freeze it, fine-tune on a sixth industrial dataset from a domain not in the paper's ablation, and compare against COCO and ImageNet-21k initializations; if the InsCore-initialized model does not match or beat those baselines on average, the domain-agnostic claim is refuted and the reported gains are benchmark-specific.","supporting_citations":[{"cited_title":"Replacing labeled real-image datasets with auto-generated contours","cited_arxiv_id":null,"evidence_quote":"This supplies the Formula-Driven Supervised Learning contour-generation method whose nested polygons InsCore turns into hollow masks."},{"cited_title":"Segrcdb: Semantic segmentation via formula-driven supervised learning","cited_arxiv_id":null,"evidence_quote":"This is the synthetic semantic-segmentation baseline and parameter source that InsCore extends to instance-level hollow masks."},{"cited_title":"Livecell—a large-scale dataset for label-free live cell segmentation","cited_arxiv_id":null,"evidence_quote":"This supplies the dense hierarchical cell-mask benchmark where InsCore shows its largest gain over COCO."},{"cited_title":"Industrial-iseg dataset, August 2024","cited_arxiv_id":null,"evidence_quote":"This supplies the manufacturing-domain benchmark of production-line defects."},{"cited_title":"Efficient load interference detection with limited labeled data","cited_arxiv_id":null,"evidence_quote":"This supplies the logistics-domain benchmark of warehouse and truck-yard objects."}],"review_version":1}