{"id":"37253d21-6523-443c-92d9-40cea175882e","arxiv_id":"1909.00182","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Scale Calibrated Training with Scale-Specific Batch Normalization lets one CNN classify images at multiple resolutions, improving ImageNet top-1 at 224 and at 128 over standard single-scale training.","lead":"This paper proposes training a single image classifier on several image sizes at once, using a batch normalization variant that keeps separate statistics for each size. It reports higher accuracy at low resolutions than standard training, at the cost of roughly M times more training compute.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Low-resolution gain is measured against a 224-trained baseline, not a 128-trained one; the paper's own ResNet-50 comparison shows +0.6 at 128, so the 11.5% headline is unsupported.","rationale":"The reader's weakest_assumption is the absence of a rule for unseen test scales, which I agree is a real limitation of S-BN (Section 3.3 switches BN parameters only for the trained scale set). However, the central claim as formulated in the paper's abstract and in the reader's strongest_claim is about the two trained scales 224 and 128, so the unseen-scale issue does not directly falsify that claim. The missing low-resolution-trained baseline does: the 11.5% number in the abstract is computed against a 224-trained network, while the paper's own Table 7 shows that against a 128-trained ResNet-50 the gain is 0.6. This means the paper's headline quantitative claim is, as stated, unsupported. The CIFAR-10 ablations with three repeated runs and the consistent S-BN improvement across ImageNet tables make the method plausible, so I would not reject; I would hold it to the same CONDITIONAL standard, with the additional explicit requirement of a 128-trained ResNet-18 baseline. Thus verdict_should_be is UNCHANGED relative to the reader's CONDITIONAL, but the condition set is sharpened.","tokens_in":9499,"tokens_out":8493,"duration_ms":77071,"concrete_test":"Train a standard ResNet-18 from scratch on ImageNet at 128x128 input resolution using the paper's reported recipe (batch size 256, 120 epochs, SGD lr 0.1 with cosine decay, standard augmentation), evaluate at 128, and repeat 3 times for a noise estimate. Compare the mean to SCT-B S-BN's 66.8@128. If the 128-trained baseline meets or exceeds 66.8, the 11.5% headline is a baseline artifact; if it is clearly below, SCT's low-resolution gain is real. Also report the same comparison for ResNet-50 using Table 7's Standard@128.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline low-resolution result is not a comparison against the correct alternative. The 11.5% figure in the abstract and Table 3 is for ResNet-18: standard BN trained at 224 gets 55.3 @128, while SCT-B with S-BN gets 66.8. But the paper's motivation is that a single SCT model replaces the need to store separate models per resolution. The relevant baseline is a network trained directly at the test resolution. The paper supplies this only for ResNet-50: Table 1 gives 73.1 for a ResNet-50 trained and tested at 128, and Table 7 lists Standard@128 = 73.3 vs SCT@128 = 73.9, a 0.6-point gain, not 11.5. No 128-trained ResNet-18 baseline is reported. If a standard ResNet-18 trained at 128 already reaches roughly 66-68, then SCT's low-resolution advantage is at best marginal and the 11.5% claim is an artifact of the mismatched 224-trained comparator. This is load-bearing for the central claim as stated: the 'improves accuracy by 11.5%' assertion is either wrong (for ResNet-50) or unverified (for ResNet-18).","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper addresses the accuracy drop that occurs when a CNN trained at a fixed resolution is evaluated at lower resolutions. The authors propose Scale Calibrated Training (SCT), in which a mini-batch is resized to several predefined scales, forwarded through shared convolutional weights, pooled to a common spatial size, and trained with a summed classification loss. To avoid what they identify as instability of vanilla batch normalization under multi-scale inputs, they propose Scale-Specific Batch Normalization (S-BN), which maintains separate running statistics and affine parameters per input scale. Experiments on ImageNet with ResNet-18/50, MobileNetV2, and quantized ResNet-18, plus CIFAR-10 ablations, are used to argue that SCT+S-BN improves accuracy at the standard 224 resolution and at lower resolutions, and that it compares favorably with Mixup, RePr, FixRes, and scale knowledge distillation.","tokens_in":9732,"tokens_out":5195,"duration_ms":104004,"significance":"If the claims were fully supported, the contribution would be practically useful: a single model that can be evaluated at several resolutions with per-scale normalization, saving storage and allowing a resolution/accuracy trade-off at inference. The paper's key technical proposal, S-BN, is simple and plausible, and the ablation study (Table 5, Figure 3) provides evidence that vanilla BN degrades under large scale gaps while per-scale BN does not. The comparison set on ImageNet, including Mixup, RePr, FixRes, and Scale KD, is reasonable. However, the headline low-resolution gain is overstated and in one form misattributed, and the method's scope is narrower than the 'no matter what sizes' claim. The current evidence establishes, at best, a moderate improvement at 128 for ResNet-50 (+0.6, Table 7) and an unverified larger gain for ResNet-18 against the wrong baseline.","major_comments":[{"comment":"The headline claim that SCT improves a single ResNet-50 by 11.5% at resolution 128 is not supported by the paper's own tables. Table 3 reports the 11.5% gain (66.8 versus 55.3) for ResNet-18, and that gain is measured against a ResNet-18 trained and tested at 224, not against a network trained directly at 128. For ResNet-50, Table 7 shows 73.9 with SCT@128 versus 73.3 for Standard@128, a 0.6-point gain. The authors should either correct the abstract and introduction or provide a ResNet-18 baseline trained and tested at 128 to substantiate the larger low-resolution advantage.","section":"Abstract; Tables 2, 3, and 7"},{"comment":"The inference procedure switches S-BN statistics and affine parameters to the scale of the input, which presupposes that every test scale belongs to the finite set of training scales described in Section 3.2. The paper's stated goal is a single network usable 'no matter what sizes of images are used for testing,' but no mechanism is given for unseen scales such as 160 or 256, and no experiment evaluates such scales. This mismatch should be resolved, either by restricting the claim to the predefined scale set or by adding an explicit rule for synthesizing or interpolating per-scale normalization statistics for unseen resolutions.","section":"Section 3.3, with reference to Sections 1 and 3.1"}],"minor_comments":[{"comment":"The text says 'we clarify the insight of our SST' and later refers to 'SST' where SCT is meant; this typo should be corrected throughout.","section":"Section 3, opening paragraph"},{"comment":"The S-BN rows use an en dash for the single-scale column; the paper should explicitly state that S-BN is defined only for multi-scale SCT training and is therefore not applicable to the single-scale baseline.","section":"Table 5"},{"comment":"The claimed distribution discrepancy between scales is supported only by a qualitative visualization; reporting a quantitative divergence measure, such as KL divergence or MMD between feature-map statistics, would make the analysis more convincing.","section":"Section 4.2, Figure 4"},{"comment":"The phrase 'catastrophic accuracy drop' overstates the numbers in Table 1, where a ResNet-50 trained at 224 still reaches 73.1% at 128; a more measured characterization of the drop would be appropriate.","section":"Section 1, Figure 1 and Table 1"},{"comment":"There are several minor language and typographical issues, including 'tge' in the contribution list and 'compacity' instead of 'capacity'; the manuscript would benefit from a careful proofread.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nHere is my take on arXiv:1909.00182. The genuinely new piece is Scale-Specific Batch Normalization (S-BN): separate BN statistics and affine parameters per training scale, switched on the fly at inference. That is not in FixRes, Mixup, or the other cited work, and the CIFAR-10 ablations support it. They run each setting three times, report means and standard deviations, and S-BN beats vanilla BN, GN, and Fixup precisely when the scale gap is large. The ImageNet gains at the standard 224 resolution are also real: ResNet-50 goes from 77.1 to 78.8, and the method carries over to MobileNetV2 and quantized ResNet-18. That is a solid empirical core.\n\nThe soft spots are mostly about presentation, but the main one is load-bearing. The abstract claims SCT improves ResNet-50 by 11.5% at size 128. That number is from Table 3, which is ResNet-18 (55.3 to 66.8). No ResNet-50 row at 128 appears in Table 3. So the abstract misattributes the result. More importantly, the low-resolution comparisons are against a 224-trained baseline, not against a network actually trained at the test scale. The paper motivates SCT as replacing separate models per resolution, so the proper baseline is a model trained at the test resolution. Table 7 supplies that comparison for ResNet-50: SCT@128 = 73.9 versus Standard@128 = 73.3. That is a 0.6-point gain, not 11.5. No 128-trained ResNet-18 baseline is reported, so the 11.5% claim is either wrong (for ResNet-50) or unverified (for ResNet-18). The honest result is still decent—one model giving 73.9 at 128 and 78.8 at 224—but it is not what the abstract promises.\n\nTwo minor issues. ImageNet numbers come without error bars, and no code or training configuration is released. Also, S-BN only works for scales seen during training; there is no interpolation rule for unseen resolutions like 160 or 256, despite the abstract saying \"no matter what sizes.\" That limitation should be stated plainly.\n\nIs the central idea still sound? Yes. The mechanism is plausible, the CIFAR-10 ablation is properly done, and the corrected comparison in Table 7 still shows SCT slightly beating a dedicated low-resolution model. The paper needs a corrected abstract, the missing 128-trained baselines, and an explicit statement about unseen scales. With those fixes, it is a contribution worth publishing.\n\nMy recommendation: send it to peer review, but expect major revision. The flaws are fixable and the core method is new enough to deserve referee time. I would bring it to the reading group as a case study in how to build fair baselines for multi-scale training, and I would cite S-BN once the numbers are cleaned up.","headline":"The 11.5% headline is misattributed and measured against the wrong baseline, but the core idea—Scale-Specific Batch Normalization—is real and worth refereeing after a fix.","tokens_in":10308,"tokens_out":3598,"would_cite":true,"duration_ms":29661,"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":"A single convolutional network can be trained to classify accurately at several image resolutions at once, using per-scale batch normalization statistics, so one model replaces many.","keywords":["Scale Calibrated Training","Scale-Specific Batch Normalization","multi-scale training","resolution discrepancy","batch normalization","image classification","low-resolution inference","ImageNet"],"falsifier":"Take an SCT-B model trained on ImageNet with scales 320, 224, 192, 128, and 64 and test it at $160\\times160$ and $256\\times256$. If accuracy at these untrained scales is no better than a standard single-scale ResNet resized to those sizes, or is far below the accuracy obtained by interpolating the running statistics of the two neighboring trained scales, then the claim that one network works for arbitrary test sizes would be disproved.","tokens_in":9233,"feed_emoji":"📏","tokens_out":8767,"duration_ms":76979,"temperature":0.7,"pith_summary":"This paper tries to establish that a single convolutional network can be trained once and then evaluated at several image resolutions, including low ones, without the large accuracy drop that normally follows when test resolution differs from training resolution. The proposed regime, Scale Calibrated Training (SCT), feeds every mini-batch to the network at multiple spatial scales, combines the classification losses, and at test time simply runs the network at whatever scale was seen in training. The paper also argues that ordinary batch normalization undermines multi-scale training because internal activation distributions drift apart across scales as the network deepens, and it introduces Scale-Specific Batch Normalization (S-BN), which keeps separate normalization statistics for each scale. On ImageNet the reported gains are 1.7% at 224x224 and 11.5% at 128x128 over standard single-scale training, making a single checkpoint usable across several resolutions and inference-speed settings.","feed_headline":"One network, many image sizes: low-res accuracy up 11.5 points","feed_subtitle":"Per-scale batch statistics let one ResNet handle several test resolutions, replacing separate models saved per size.","key_machinery":"The load-bearing object is Scale-Specific Batch Normalization (S-BN), a normalization layer that keeps a separate tuple of running mean, running variance, and affine parameters $(\\mu^i, \\sigma^i, \\gamma^i, \\beta^i)$ for each input scale $s_i$ used in training. In the SCT forward pass, a mini-batch is resized to $M$ predefined scales, all copies go through the same convolutional layers, and the loss is a weighted sum over scales; at inference the network switches to the BN statistics of the input's scale. This mechanism isolates the scale-dependent part of the internal feature distribution, which the paper shows diverges across scales in deeper layers, so the shared convolutional weights can be trained without a single normalization layer trying to fit contradictory statistics.","core_discovery":"The central discovery is that the failure of naive multi-scale training is largely a normalization failure, not a capacity failure. When a single network is trained on inputs of different resolutions, the per-channel mean and variance of intermediate activations diverge between scales as depth increases; a shared batch-normalization layer then tries to normalize all scales with one set of statistics and hurts accuracy. Scale-Specific Batch Normalization gives each scale its own running mean, running variance, and affine parameters, selected on the fly at inference, and restores the multi-scale gains. The result is a single model whose accuracy at a trained low resolution approaches that of a model trained specifically at that resolution, while keeping full-resolution accuracy.","pith_inferences":["Editorial inference: if S-BN's advantage comes from separating per-scale statistics rather than from the multi-scale data augmentation per se, the same per-scale normalization idea should transfer to other settings where input statistics shift systematically, such as different aspect ratios or object scales in detection; the paper does not test this.","Editorial inference: the paper leaves open how to handle a resolution between trained scales; interpolating the running means and variances of neighboring S-BN layers is a natural fix, and its success would be a direct test of whether the per-scale statistics vary smoothly with resolution.","Editorial inference: the 11.5-point low-resolution gain is reported for a model also trained at high resolutions, so SCT may be understood as a form of resolution-aware data augmentation plus normalization adaptation, which likely benefits any architecture whose batch statistics shift with input size."],"forward_implications":["One checkpoint can serve multiple resolutions, so deploying several sizes no longer requires K separate models or per-size storage.","Low-resolution inference becomes practical on the same network: the reported ResNet-18 accuracy at 128x128 rises from 55.3% to 66.8% with SCT-B.","Standard-resolution accuracy is not traded away: ImageNet top-1 at 224x224 improves over the single-scale baseline, with the paper's headline result reporting a 1.7% gain for ResNet-50.","The method is architecture-agnostic enough to help limited-capacity models: MobileNetV2 and 8-bit quantized ResNet-18 both gain accuracy.","Because S-BN only swaps normalization statistics at test time, the multi-scale capability adds no inference cost."],"supporting_citations":[{"why":"Defines batch normalization, the module whose shared-statistics failure motivates S-BN.","marker":"[9]"},{"why":"Supplies the ResNet architectures used in the main ImageNet experiments and ablations.","marker":"[5]"},{"why":"Provides the main resolution-discrepancy baseline that SCT is compared against.","marker":"[19]"},{"why":"Serves as a training-scheme baseline whose accuracy SCT must beat.","marker":"[25]"},{"why":"Serves as a training-scheme baseline based on channel re-pruning, compared in the ImageNet table.","marker":"[13]"},{"why":"Supplies group normalization as an alternative normalization used in the ablation study.","marker":"[23]"},{"why":"Supplies Fixup initialization as a normalization-free baseline in the ablation.","marker":"[26]"},{"why":"Supplies the PACT quantization method used in the limited-capacity 8-bit ResNet experiments.","marker":"[2]"},{"why":"Supplies the MobileNetV2 architecture used to test the method on limited-capacity models.","marker":"[14]"}],"fun_headline_variants":["Per-scale batch stats lift low-res accuracy 11.5 points","Multi-scale training fixed by scale-specific normalization","One ResNet, many resolutions: per-scale stats win","Low-res accuracy up 11.5 points with scale-calibrated training","Scale-specific normalization solves multi-scale training"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes the list of test image sizes is fixed in advance and that every test image is one of those exact sizes, because the network keeps separate normalization statistics only for the scales it trained on.","fun_headline_variants_meta":{"raw":{"variants":["Per-scale batch stats lift low-res accuracy 11.5 points","Multi-scale training fixed by scale-specific normalization","One ResNet, many resolutions: per-scale stats win","Low-res accuracy up 11.5 points with scale-calibrated training","Scale-specific normalization solves multi-scale training"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000141,"raw_usage":{"total_tokens":1122,"prompt_tokens":858,"completion_tokens":264,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":474,"completion_tokens_details":{"reasoning_tokens":184}},"tokens_in":474,"tokens_out":264,"duration_ms":3400,"temperature":1.0,"reasoning_tokens":184,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T05:58:47.273058+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take an SCT-B model trained on ImageNet with scales 320, 224, 192, 128, and 64 and test it at $160\\times160$ and $256\\times256$. If accuracy at these untrained scales is no better than a standard single-scale ResNet resized to those sizes, or is far below the accuracy obtained by interpolating the running statistics of the two neighboring trained scales, then the claim that one network works for arbitrary test sizes would be disproved.","supporting_citations":[{"cited_title":"Repr: Improved training of convolutional filters","cited_arxiv_id":null,"evidence_quote":"Serves as a training-scheme baseline based on channel re-pruning, compared in the ImageNet table."},{"cited_title":"Group normalization","cited_arxiv_id":null,"evidence_quote":"Supplies group normalization as an alternative normalization used in the ablation study."}],"review_version":1}