{"id":"67e1611f-e910-4b99-97f8-e3cbf20ac2b5","arxiv_id":"1908.07323","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"ISN filters extreme-scale objects during multi-scale training and testing, improving COCO object detection, instance segmentation, and human pose estimation.","lead":"Instance Scale Normalization (ISN) is a training and testing strategy that ignores objects whose size falls outside a fixed scale range during multi-scale object detection. This simple filter consistently improves accuracy on COCO detection, segmentation, and human pose estimation across multiple backbones.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ISN's reported gains may be due to SNIPER-style chip sampling, not the [16,560] scale filter: no ablation separates chips from scale normalization in the central comparisons.","rationale":"The reader's weakest assumption was that the validation-tuned ISN range [16, 560] might overfit. My concern is different and more load-bearing: the experimental protocol conflates the proposed scale-range filtering with SNIPER-style chip sampling, so the attribution of the headline improvements to 'instance scale normalization' is not yet established. The paper does have real strengths: consistent gains across detection, segmentation, and keypoint tasks; multiple backbones; a shared codebase for baseline and ISN; controlled use of DCN and Soft-NMS; and a strong test-dev number. These make the empirical phenomenon credible. However, because ISN explicitly inherits SNIPER's chip generation and the baselines do not, the reported +2 to +5 AP gains could largely reflect the known benefits of focusing training on object-dense chips rather than the uniform scale range. This is an internal-validity issue, not a disagreement with the community's consensus about scale variation. A single two-arm ablation would settle it: add chips to the multi-scale baseline without the [16, 560] filter, or remove chips from ISN. If the ablation shows the filter still matters, the central claim stands; if not, the paper should be reframed as combining chip sampling with a consistent scale range. The reader's conditional verdict remains appropriate, but the condition should explicitly include this ablation rather than only robustness of the tuned range.","tokens_in":12905,"tokens_out":10907,"duration_ms":119424,"concrete_test":"Re-run the Table 4 ResNet-50-FPN comparison with an additional arm: the same multi-scale train/test baseline plus SNIPER-style chips (same patch size 576, same scaling factors {4, 2, 1, 0.5, 0.25}, same codebase and Soft-NMS) but without the [16, 560] filtering. If this arm matches ISN's validation AP (42.9, or 45.0 with DCN), then the scale-range filtering contributes little beyond chip sampling. If ISN still leads by about 2 AP, the concern is resolved.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central comparisons in Tables 4, 6, 8, and 9 compare a composite pipeline against a plain multi-scale baseline, so the reported gains cannot be cleanly attributed to instance scale normalization. Section 3.3.1 states that ISN uses 'the same chip generation strategy in SNIPER [27]', and Table 7 refers to 'the patch size 576 used in ISN training'; Section 4.3.2 likewise mentions 'patches used in ISN training here'. The multi-scale baseline, by contrast, follows Mask R-CNN-style full-image multi-scale training and testing with no chip sampling (Section 4.1). Because SNIPER already established that object-centric chip sampling improves detection, the +2 to +5 AP improvements reported for ISN may come from this inherited component rather than from compressing object scales into [16, 560]. The paper provides no ablation that removes chips from ISN or adds chips to the baseline, so the specific causal claim that uniform scale-range filtering drives the improvement is currently uncontrolled. This is a more direct threat to the central claim than the secondary concern about greedy tuning of the range on the validation set.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Instance Scale Normalization (ISN), a training and testing scheme that restricts object instances to a consistent scale range [sl, su] (set to [16, 560]) across all image-pyramid resolutions. Using the same chip-generation strategy as SNIPER, ISN discards objects outside the range in both training and testing, with the goal of compressing the scale space and reducing optimization difficulty. Experiments on COCO report consistent improvements over multi-scale training/testing baselines on object detection (Tables 4, 6), instance segmentation (Table 8), and human pose estimation (Table 9), with a single ResNet-101 model reaching 46.5 mAP on test-dev (Table 3).","tokens_in":13111,"tokens_out":2790,"duration_ms":29772,"significance":"If the improvements are attributable to ISN's scale-range filtering, the method would offer a simple, architecture-agnostic alternative to multi-scale training and testing, with practical benefits for small backbones and single-scale inference. The paper is honest about the validation-set greedy search for the ISN range (Section 4.2.1) and provides consistent, controlled baselines across three tasks and several backbones, which is a strength. The central empirical claim is plausible, but the paper currently does not separate the contribution of scale-range filtering from the contribution of the inherited SNIPER-style chip sampling, and the SOTA comparison in Table 3 is confounded by the use of FPN in ISN but not in SNIP/SNIPER. These issues require additional experiments before the claimed attribution can be accepted.","major_comments":[{"comment":"The central comparisons do not isolate the effect of scale-range normalization from the effect of SNIPER-style chip sampling. Section 3.3.1 states that ISN uses 'the same chip generation strategy in SNIPER [27]', and Sections 4.2.4 and 4.3.2 refer to the patch size 576 and patches used in ISN training. The multi-scale baselines, by contrast, are trained on full images with image pyramids and no chip sampling. Since SNIPER already demonstrated that object-centric chip sampling improves detection, the reported +2 to +5 AP gains may be largely due to the inherited chip mechanism rather than the [16, 560] scale filter. An ablation is needed: e.g., (a) add SNIPER-style chips to the multi-scale baseline without scale filtering, and (b) run ISN on full images without chips, while keeping all other settings fixed. Without such an ablation, the causal claim that compressing the scale space drives the improvement is uncontrolled.","section":"Section 3.3.1; Tables 4, 6, 8, 9"},{"comment":"The state-of-the-art comparison is confounded by FPN. SNIP and SNIPER are reported on ResNet-50/ResNet-101 without FPN, while ISN is reported on ResNet-*-FPN backbones. The text claims ISN 'surpasses previous methods hugely', but the comparison mixes a change in scale-sampling method with a change in network architecture. A fair comparison would include SNIP or SNIPER augmented with FPN, or ISN without FPN, or otherwise hold the feature-pyramid component fixed. This is important because the paper's own Section 3.3.2 argues that FPN integration is a deliberate and beneficial part of ISN; the reader cannot tell how much of the Table 3 margin comes from ISN versus from FPN.","section":"Table 3 (SOTA comparison)"},{"comment":"The ISN range [16, 560] is selected by greedy search on the COCO validation set, and the same validation set is used to report most of the subsequent improvements (Tables 4, 6, 8, 9). This creates a risk of selection bias in the reported margins, since the hyperparameter is optimized on the evaluation set. The paper would be strengthened by reporting results on the validation set under a pre-registered default range (e.g., [16, 640] or [0, 560]) or by reporting test-dev results for all key configurations. The single test-dev result in Table 3 is welcome, but it is not accompanied by an ablation that removes the chip confound, so it does not resolve the attribution issue in the first major comment.","section":"Section 4.2.1 and Tables 4-9"}],"minor_comments":[{"comment":"There is a spelling error: 'normalizatoin' should be 'normalization'.","section":"Section 3.2.1"},{"comment":"The description of the third greedy-search step reads 'sl was set to 560 and su = 32 is evaluated continuously'. This appears to be a typo: the context indicates sl=32 and su=560 should be evaluated. Please correct.","section":"Section 4.2.1"},{"comment":"The row label 'ResNet-101-FP' should read 'ResNet-101-FPN'. Also, the footnote marks SNIPER as evaluated on COCO test-dev, while the '+ISN' row has no asterisk; if the ISN row is on COCO val2017, the comparison between SNIPER and +ISN within this table mixes evaluation sets and should be clarified or moved to Table 3.","section":"Table 4"},{"comment":"The claim that ISN is 'robust for scale variation' when changing from a predefined test resolution to the original resolution is based on a single set of numbers in Table 7; reporting the standard deviation across multiple runs or the per-scale breakdown would make the robustness claim more convincing.","section":"Section 4.2.4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is simple and potentially useful, and the empirical work is extensive. However, the missing ablation separating chip sampling from scale normalization is a substantive gap that should be addressed before the paper can be considered for acceptance. The Table 3 FPN confound also needs to be fixed. I would not reject the manuscript: the concern is addressable with additional experiments within the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, this is a real extension of SNIP/SNIPER: a single [sl, su] scale range applied consistently across all pyramid resolutions and across train and test, replacing SNIP's per-resolution ranges. Second, the empirical gains in the headline tables are probably not purely due to that scale filter, because the method inherits SNIPER's chip sampling and the paper never ablates chips away.\n\nWhat is genuinely good: the paper is honest about its lineage. It cites SNIP and SNIPER, explains SNIP's inconsistent selection with a nice figure, and shows consistent AP improvements across detection, segmentation, and keypoint tasks on several backbones, with small-object AP especially improved. The greedy validation tuning of [16, 560] is reported openly, not hidden. That is more than many vision papers do.\n\nWhere it gets soft: the stress-test is correct on the main empirical question. The baseline is multi-scale full-image training; ISN is chips + consistent range + FPN. Chips are known to help (SNIPER), so the +2-5 AP cannot be assigned to 'instance scale normalization' specifically. Table 3 also compares ISN (+FPN) to SNIPER (no FPN), which is not a controlled comparison. A single ablation adding chips to the MST baseline or removing chips from ISN would settle it; the paper doesn't provide one. The greedy range search is a second, lesser concern, and with no released code or seeds, the exact numbers are hard to verify. None of this makes the results obviously wrong; it makes the paper's central causal claim uncontrolled.\n\nWho this is for: people working on scale-invariant training recipes. There is a useful idea here worth debating, and the empirical evidence, even if confounded, is not noise. I would send it to a referee, but the referee should demand the chip ablation. If the authors supply it and the gains hold, this becomes a solid paper. Without it, treat the specific numbers as a composite recipe.\n\nMy recommendation: engage with the work, but do not accept the attribution as established.","headline":"A real but under-ablated extension of SNIP/SNIPER: the consistent scale range is a genuine idea, but the reported gains are confounded by inherited chip sampling.","tokens_in":13645,"tokens_out":2954,"would_cite":true,"duration_ms":28752,"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":"Restricting every object to a single size range [16,560] during both training and testing compresses the scale space of objects and improves detection, segmentation, and keypoint estimation over multi-scale baselines.","keywords":["instance scale normalization","object detection","scale variation","multi-scale training","feature pyramid network","COCO","instance segmentation","human pose estimation"],"falsifier":"Train an ISN detector on COCO with the range [16,560] and evaluate it on a dataset whose object scales are mostly below 16 pixels or above 560 pixels, such as aerial or pathology images; if the method's accuracy falls below that of a standard multi-scale detector, it would show that the gains come from matching the tuned range rather than from the normalization principle itself.","tokens_in":12690,"feed_emoji":"📏","tokens_out":4096,"duration_ms":35502,"temperature":0.7,"pith_summary":"This paper claims that the longstanding difficulty of detecting objects at very different sizes can be largely neutralized by forcing all objects into a single, fixed size range during both training and testing. The method, called Instance Scale Normalization (ISN), resizes each image through a fixed set of scales and then keeps only those objects whose size in the resized image falls between 16 and 560 pixels. ISN does not change the network architecture, yet the paper reports that it consistently outperforms standard multi-scale training and testing on object detection, instance segmentation, and human pose estimation across several backbones, including 46.5 mAP on COCO test-dev with a ResNet-101. The paper presents this as evidence that scale variation is not a property the network must learn to absorb, but a range of inputs that can be filtered.","feed_headline":"One fixed size range beats multi-scale detection","feed_subtitle":"ISN filters every object to 16–560 pixels and beats multi-scale training on COCO by up to 5 mAP.","key_machinery":"The load-bearing device is the ISN range [sl, su] = [16, 560] applied as a filter on object scales in resized images. It converts scale variation from a continuous wide spectrum into a bounded interval, and the same interval is reused for training and testing, so the distribution shift between phases is removed. The range is found by a greedy search on the COCO validation set: first the lower bound is raised from 0 to 16, then the upper bound is lowered from 640 to 560, each step chosen by local mAP. With FPN integrated, each pyramid stage receives objects in a narrower effective range, which the paper argues eases optimization and avoids the self-contradictory sampling it attributes to SNIP.","core_discovery":"ISN's central claim is that a consistent scale interval applied identically in training and testing compresses the scale space of objects, letting a detector work in a moderate, uniform regime instead of facing extreme sizes. For a multi-resolution image pyramid with scaling factors {4.0, 2.0, 1.0, 0.5, 0.25}, ISN first resizes each image, then discards any proposal whose scale in that resized image lies outside [16,560]. The same rule is used at test time, so the model never sees objects outside this range in either phase. Because feature pyramid network (FPN) stages each specialize in a sub-range of scales, ISN integrates with FPN to keep a wide feasible range while still letting each stage train on objects of consistent size. The paper reports that this yields gains over multi-scale training and testing baselines in all three tasks, with the largest relative gains on small objects and on keypoint detection.","pith_inferences":["If the mechanism is genuinely about compressing scale space rather than about the specific [16,560] bound, the same trick may apply to other continuous nuisance variations—such as rotation, aspect ratio, or illumination—where extreme values are rare but harmful.","Because the range is tuned on COCO val, a stress test on a dataset with a different scale distribution (for example, aerial imagery with mostly small objects) would show whether the gains come from the principle itself or from matching the evaluation range.","The reported success suggests a testable extension: instead of a fixed range, one could learn a per-dataset or per-class range that adapts to the object scale distribution, which might outperform the hand-tuned interval."],"forward_implications":["A detector trained with ISN works on a specified scale range [16,560] and can be tested at a single resolution, giving accuracy nearly as high as multi-scale testing while running faster.","ISN improves performance on small objects substantially; for example, APs on COCO val with ResNet-50-FPN rises from 26.9 with multi-scale training to 31.4 with ISN.","ISN transfers beyond detection: it improves instance segmentation mask AP by around 1 point and keypoint AP by more than 3.5 points over multi-scale training and testing.","ISN is backbone-agnostic, bringing the largest relative gains to small models like ResNet-18 and MobileNet-v2, making it a cheap accuracy boost for real-time applications.","The single ResNet-101 ISN model reaches 46.5 mAP on COCO test-dev, competitive with state-of-the-art detectors that use heavier machinery."],"supporting_citations":[{"why":"SNIP is the scale-invariance baseline ISN compares against, and its inconsistent sampling is the specific defect ISN claims to fix.","marker":"[26]"},{"why":"SNIPER provides the efficient multi-scale training baseline and also the chip-generation strategy that ISN adapts for its scale normalization.","marker":"[27]"},{"why":"FPN is the feature pyramid network that ISN integrates with to maintain a wide feasible scale range while each stage handles a smaller sub-range.","marker":"[15]"},{"why":"Mask R-CNN is the unified framework on which ISN is evaluated for instance segmentation and human pose estimation.","marker":"[11]"},{"why":"Faster R-CNN is the detection framework on which ISN is implemented and compared against multi-scale baselines.","marker":"[24]"},{"why":"The COCO dataset provides the object scale distribution that motivates the ISN range and supplies the benchmarks for all experiments.","marker":"[17]"}],"fun_headline_variants":["ISN normalizes instance scale to a fixed range","Instance Scale Normalization beats multi-scale detection","One scale range for all objects: ISN wins on COCO","Scale variation tamed by instance-level normalization","ISN compresses scale space, boosting detection and pose"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The fixed ISN range [16,560] is chosen by a greedy search on the COCO validation set, and the paper assumes that this range transfers to the test distribution and to other tasks and backbones without re-tuning, so the reported gains may depend on matching the validation distribution.","fun_headline_variants_meta":{"raw":{"variants":["ISN normalizes instance scale to a fixed range","Instance Scale Normalization beats multi-scale detection","One scale range for all objects: ISN wins on COCO","Scale variation tamed by instance-level normalization","ISN compresses scale space, boosting detection and pose"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1339,"prompt_tokens":912,"completion_tokens":427,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":528,"completion_tokens_details":{"reasoning_tokens":351}},"tokens_in":528,"tokens_out":427,"duration_ms":4978,"temperature":1.0,"reasoning_tokens":351,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T12:19:46.919140+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train an ISN detector on COCO with the range [16,560] and evaluate it on a dataset whose object scales are mostly below 16 pixels or above 560 pixels, such as aerial or pathology images; if the method's accuracy falls below that of a standard multi-scale detector, it would show that the gains come from matching the tuned range rather than from the normalization principle itself.","supporting_citations":[{"cited_title":"Singh and L","cited_arxiv_id":null,"evidence_quote":"SNIP is the scale-invariance baseline ISN compares against, and its inconsistent sampling is the specific defect ISN claims to fix."},{"cited_title":"Singh, M","cited_arxiv_id":null,"evidence_quote":"SNIPER provides the efficient multi-scale training baseline and also the chip-generation strategy that ISN adapts for its scale normalization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FPN is the feature pyramid network that ISN integrates with to maintain a wide feasible scale range while each stage handles a smaller sub-range."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mask R-CNN is the unified framework on which ISN is evaluated for instance segmentation and human pose estimation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Faster R-CNN is the detection framework on which ISN is implemented and compared against multi-scale baselines."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The COCO dataset provides the object scale distribution that motivates the ISN range and supplies the benchmarks for all experiments."}],"review_version":1}