{"id":"798fde3a-a983-4c8e-b919-dcfd6e5f3c07","arxiv_id":"2501.01733","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A plug-and-play augmentation that mixes same-label item patches and suppresses mismatched high-IoU losses improves X-ray prohibited-item detection under category and bounding-box annotation noise.","lead":"This paper proposes Mix-Paste, a data augmentation method that blends multiple X-ray item patches sharing the same category label and pastes the mixture back into the training image, together with a loss-suppression strategy called LLS, to make prohibited-item detectors robust to noisy annotations. It reports large detection gains on noisy X-ray benchmarks and shows the approach also helps on noisy MS-COCO.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Mix-Paste's probability argument assumes random label noise, but the paper's own noise analysis describes visually similar categories confused in a structured way; the high-noise gains may not transfer to real annotation noise.","rationale":"The reader's weakest assumption identifies the synthetic noise injection protocol as the main threat to generality. My stress-test agrees and sharpens it: the paper itself describes the real noise as confusion among visually similar items, which violates the independence/randomness assumption behind the 1 - Pc^K argument. This is not a disagreement with community consensus but a correctness risk in transferring the empirical result. The paper does provide useful independent support: released code, an original-OPIXray experiment showing gains even without injected noise, ablations across detectors and noise distributions, and a consistent pattern across OPIXray, PIDray, and COCO. Those make the method plausible, but they do not establish that the large high-noise gains would survive structured noise. The reader's CONDITIONAL verdict remains appropriate: acceptance should be contingent on a structured-noise evaluation or a direct natural-noise benchmark. No verdict change is needed, hence UNCHANGED.","tokens_in":27326,"tokens_out":7780,"duration_ms":91068,"concrete_test":"Repeat the OPIXray experiments at Pc=60%, Pb=60% (Table I) replacing uniform random label flips with a structured transition matrix that corrupts only visually similar classes (e.g., using empirically observed confusions among folding/utility/straight knives), keeping K=2, p=0.6, and all other settings fixed. If Mix-Paste+LLS's margin over FRCNN falls to less than half of the reported 25.1 mAP@.5 gain, the central claim of robustness to realistic noisy annotations is unsupported. A complementary check is to re-annotate a few hundred OPIXray training boxes to estimate the true label confusion matrix and train directly on those naturally noisy annotations.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central theoretical justification in Section III-C, 1 - Pc^K, requires that same-label patches are independently and randomly corrupted, so that mixing K patches almost surely includes at least one true instance. But Section IV-B states that OPIXray's natural category noise comes from the great similarity between certain prohibited items (folding/utility/straight knives). Under such structured confusion, labels are not flipped uniformly at random: a true 'utility knife' may almost always be annotated as 'folding knife', so patches sharing the noisy label can contain no true same-label item even for arbitrarily large K. In that regime, Mix-Paste does not increase the probability of a correct label and may instead amplify the wrong visual pattern. LLS also behaves differently under structured noise: it suppresses losses for any non-background prediction differing from the noisy label, which includes the model's correct prediction of the confusable true class, while predictions matching the noisy label are still trained as positive. The reported 25.1-point mAP@.5 gain at Pc=60%, Pb=60% comes from the synthetic random-noise protocol in Section IV-C; the only natural-noise evidence is the low-noise (~5%/4%) Table VII result, which cannot validate the high-noise claim. Therefore the load-bearing assumption is that random uniform label replacement reproduces real X-ray annotation noise, and the paper itself provides evidence that it does not.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Mix-Paste, a data augmentation method for training X-ray prohibited-item detectors under noisy annotations (category label noise and bounding-box noise). For each ground-truth item patch, Mix-Paste randomly selects K same-label patches from other images, blends them via an edge-smoothing mask, and pastes the blend back into the original bounding box. The authors argue that mixing increases the probability that the augmented patch contains a correctly labeled item (1 - P_c^K) and, analogously, that it improves box accuracy (1 - P_b^K). They also introduce an item-based large-loss suppression (LLS) rule that discards classification losses for high-IoU predictions whose predicted label differs from the ground-truth label. Experiments on OPIXray, PIDray, and noisy MS-COCO show that Mix-Paste plus LLS substantially improves mAP over Faster R-CNN and several baselines under synthetically injected uniform random noise, with a modest gain on the original (naturally noisy) OPIXray dataset.","tokens_in":27642,"tokens_out":7262,"duration_ms":71642,"significance":"If the claims hold, the paper introduces a simple, plug-and-play augmentation direction for noisy-annotation object detection, which is a practical and under-explored problem for X-ray security imagery. The strengths are the breadth of experiments (two X-ray datasets plus MS-COCO, multiple noise rates, ablations of components, and comparison with several specialized baselines), the release of source code, and the clear presentation of the method. The main limitation is that the high-noise evidence is generated entirely from a uniform random synthetic noise protocol, whereas the paper's own characterization of real X-ray noise is structured (confusions among visually similar knife types). The significance is therefore real but currently contingent on the validity of the random-noise assumption.","major_comments":[{"comment":"The central probability argument, 1 - P_c^K, assumes that each same-label patch is independently corrupted with probability P_c under a random label-noise model. However, Section IV-B states that real category noise in OPIXray arises from the visual similarity between knife types (folding, utility, straight, multi-tool), i.e., a structured confusion pattern. Under such structured noise, K patches sharing the same noisy label may all be the same wrong class, so mixing does not increase the chance of including a true same-label item. All high-noise results (Tables I, II, IV) use the synthetic uniform random noise of Section IV-C; the only natural-noise evidence (Table VII) is at an estimated ~5% category noise, which cannot validate the high-noise claim. The authors should either test Mix-Paste under a structured noise model (e.g., a confusion matrix biased by visual similarity) or clearly scope the claims to random label noise.","section":"Section III-C, Eq. (1); Section IV-B; Section IV-C"},{"comment":"The claimed bounding-box benefit, 1 - P_b^K, is not supported by the algorithm as described. The mixed patch is pasted into the original (possibly noisy) bounding box, and the ground-truth box coordinates are unchanged, so mixing K patches does not by itself increase the probability that the box 'accurately bounds a correct prohibited item.' The empirical results in Fig. 5(b) suggest Mix-Paste helps under box-noise-only conditions, but the stated mechanism is not convincing. The theoretical explanation should be revised to describe how the augmentation actually interacts with box noise, and an ablation that isolates box noise (e.g., with clean labels) should be reported at the same noise rates as in the main tables.","section":"Section III-C, Eqs. (1)-(3)"},{"comment":"LLS suppresses all non-background predictions whose label differs from the ground-truth label at high IoU, not only predictions corresponding to extra items introduced by Mix-Paste. In the very scenario addressed by the paper, the ground-truth label itself may be wrong; in that case the model's correct prediction of the true class would be suppressed, effectively training the model toward the noisy label. The paper demonstrates the benefit of LLS under random synthetic noise (Table XIV) but does not analyze this interaction with noisy ground-truth labels. A more selective suppression rule, or an experiment that separates LLS's effect on mixed-patch extra items from its effect on single-patch noisy labels, would strengthen the claim.","section":"Section III-D, Eq. (4)"}],"minor_comments":[{"comment":"The definition of the edge-smoothing mask appears reversed: the formula assigns α ≈ 1 near the patch boundary (small d) and a constant λ in the interior, which would make the boundary dominated by the base patch rather than smoothing the edge. Please verify the formula or its verbal description; the ablation in Table VI suggests the mask has a minor effect, but the current definition is confusing.","section":"Eq. (2)"},{"comment":"Several numbers in Table II have missing plus signs (e.g., '21.5 (1.6)' and '11.1(-2.2)'); the sign convention should be made consistent with Table I.","section":"Tables I-II"},{"comment":"The conclusion refers to an 'LSS strategy'; the abbreviation should be 'LLS' for consistency with Section III-D.","section":"Conclusion"},{"comment":"The reference list appears twice in the arXiv version (pages 14-16); the final version should contain a single bibliography.","section":"References"},{"comment":"The complete collapse at p=1 (0.4% mAP) is a striking failure mode; a one- or two-sentence explanation beyond the train/test distribution mismatch would help readers understand the sensitivity to this hyperparameter.","section":"Table X"}],"recommendation":"major_revision","confidential_remarks":"The paper is well written and the method is simple and potentially useful, but the main risk is that the high-noise experiments all use uniform random synthetic noise, which the paper's own noise analysis suggests is not the structure of real X-ray annotation errors. I would not reject because the idea is novel and the low-noise real-data result is encouraging, but I would ask for either a structured-noise experiment (e.g., biased confusion matrix) or a clear scope restriction before publication. The bounding-box probability argument also needs to be fixed or removed, as it is currently not aligned with the algorithm."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Mix-Paste is a genuine new combination: mixing several patches that share the same annotated label and pasting the mixture back into the original box, plus a large-loss suppression rule for the extra items the mixing can create. The parts are close to CutMix/Mosaic and existing loss-suppression ideas, but the specific pairing and the motivation (increasing the chance that a noisy box contains the true item) are new. The paper also does a lot of things right empirically: three datasets, several noise rates, ablations on K, p, edge smoothing, the mixing strategy, and comparisons to both detection baselines and label-noise methods. The ablation where random cross-category mixing fails while same-label mixing works is informative, as is the comparison against three versions of the small-loss criterion. Code is released.\n\nNow the soft spots. First, all numbers come from single runs; no error bars or significance tests. Given the large reported gaps at high noise, that is a minor issue, but it would be nice. Second, K and p are chosen via ablations on the same dataset, so there is some risk of selection on the test set. Also minor.\n\nThe larger concern is external validity. The probability argument 1 - Pc^K assumes that corrupted labels are flipped uniformly at random. But the paper's own noise analysis for OPIXray says category errors come from great similarity between folding/utility/straight knives. Under that kind of structured confusion, patches sharing a noisy label may all be wrong in the same way; mixing K of them does not increase the chance of a correct label, and LLS may end up suppressing the model's correct prediction while training on the wrong label. The 25-point gain at 60%/60% comes from the synthetic random-noise protocol. The only natural-noise evidence is the low-noise original dataset (~5% category, ~4% box), which cannot justify the high-noise claim. This is not a fatal flaw, but it is a real limitation. The paper would be much stronger with a structured-noise experiment (e.g., an asymmetric confusion matrix among visually similar classes) or a larger naturally noisy validation set.\n\nOverall, the central claim—that augmentation can be an inference-free tool against noisy detection annotations—holds up under the tested conditions. I would send this to a serious referee, and I'd ask them to push for structured-noise validation and error bars.","headline":"A plausible and well-tested augmentation method for noisy X-ray detection, but the headline numbers are all under synthetic random noise; the structured-noise gap is real and should be addressed before publication.","tokens_in":28152,"tokens_out":3332,"would_cite":false,"duration_ms":33190,"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":"The paper shows that blending same-label item patches into a mixed patch and pasting it back makes X-ray prohibited-item detectors robust to both category-label and bounding-box noise.","keywords":["X-ray prohibited item detection","noisy annotations","data augmentation","Mix-Paste","large-loss suppression","object detection","category label noise","bounding box noise"],"falsifier":"Re-annotate a subset of OPIXray or PIDray with clean labels, inject structured noise such as confusion concentrated between visually similar categories and boxes consistently inflated in one direction, and check whether Mix-Paste plus LLS keeps its advantage; if the gain collapses, the independence assumption behind $1-P_c^K$ is the weak link.","tokens_in":27160,"feed_emoji":"🛡️","tokens_out":8989,"duration_ms":81772,"temperature":0.7,"pith_summary":"This paper argues that data augmentation can absorb a large share of annotation noise in X-ray prohibited-item detection, even when both category labels and bounding boxes are corrupted. The proposed Mix-Paste method mixes K item patches that carry the same category label and pastes the blended patch back into the image, which raises the chance that the boxed region really contains an item of that class. A companion item-based large-loss suppression (LLS) rule drops the classification losses of plausible detections whose predicted category contradicts the noisy ground-truth label. On OPIXray with 60% category noise and 60% box noise, the combination lifts mAP@.5 from 56.7% to 81.8% and mAP@[.5,.95] from 18.4% to 33.7%, and similar gains appear on PIDray and, more modestly, on noisy MS-COCO. The paper's central claim is that this inference-free, plug-and-play recipe is a general way to train robust detectors without needing clean data or noise-rate estimates.","feed_headline":"Mix same-label patches; detectors hold under 60% noisy X-ray labels","feed_subtitle":"Blending K same-label item patches makes it likelier a true item survives, beating the baseline by 25 points at 60% noise.","key_machinery":"The central object is the mixed patch $\\hat{B} = \\alpha \\odot B_a + \\sum_{n=2}^K \\frac{1-\\alpha}{K-1} \\odot \\mathrm{resize}(B_n)$, where $B_a$ is the original patch, the $B_n$ are randomly chosen patches with the same category label, and $\\alpha$ is an edge-smoothing mask. This patch carries the probability amplification $1-P_c^K$ and is pasted back at the original box location. The second mechanism is item-based LLS, which splits predictions into negative, background, positive, and potentially-positive categories and discards the classification loss of the potentially-positive ones (high IoU, non-background label differing from ground truth). Together the machinery converts noisy labels into a training signal dominated by correct items while suppressing the loss spikes caused by extra blended items.","core_discovery":"Under noisy annotations, a single ground-truth box with label $c$ contains a correctly labeled item with probability roughly $1-P_c$ when the category noise rate is $P_c$. If K patches all labeled $c$ are mixed, the probability that the resulting region does not contain a correct item of class $c$ falls to $P_c^K$, so the chance the region is genuinely correct rises to $1-P_c^K$; the same reasoning applies to bounding-box noise with rate $P_b$, giving $1-P_b^K$. Mix-Paste implements this by resizing K same-label patches to one size, linearly blending them with an edge-smoothing mask, and pasting the result over the original box. Because the blended patch can also introduce extra items whose labels are not recorded, LLS identifies high-IoU predictions whose label disagrees with the ground-truth label and excludes their classification losses from the training objective. The paper reports that this two-part recipe consistently outperforms prior robust-detection and label-noise methods on OPIXray and PIDray across noise rates from 20% to 60%, and transfers to MS-COCO.","pith_inferences":["Beyond the paper's random-noise protocol, the probability gain $1-P_c^K$ presumes that label mistakes are independent; under systematic confusion of visually similar knife types, mixing may consolidate the wrong class rather than dilute it.","A natural next experiment would separate Mix-Paste's noise-suppression effect from its overlap-mimicking effect by running it on a fully clean, non-overlapping dataset; the original-OPIXray result suggests both effects are present.","Because Mix-Paste is a data-level operation, it should combine with label-refinement and loss-correction methods; the paper does not test such a combination."],"forward_implications":["At 60% category noise and 60% box noise on OPIXray, Mix-Paste with LLS raises mAP@.5 from 56.7% to 81.8% and mAP@[.5,.95] from 18.4% to 33.7%, a gain of 25.1 and 15.3 points over the Faster R-CNN baseline.","Similar gains hold on PIDray, where the method improves average mAP@[.5,.95] by 17.0 points at the 60% noise level across easy, hard, and hidden test splits.","The recipe is plug-and-play: applying Mix-Paste alone improves two-stage, one-stage, and X-ray-specific detectors, and adding LLS gives a further gain over Mix-Paste alone.","On noisy MS-COCO, the same augmentation improves over the baseline by about 2.1 mAP@.5 and 2.9 mAP@[.5,.95] at 60% noise, showing the idea is not restricted to X-ray imagery.","The paper also reports that mixing with large K (4 or 5) or applying Mix-Paste to essentially all training images hurts performance; K=2 or 3 and a 0.6 mixing probability are the stable choices."],"supporting_citations":[{"why":"Supplies the OPIXray benchmark and the real noisy-annotation examples that motivate the method.","marker":"[1]"},{"why":"Supplies the PIDray benchmark with easy, hard, and hidden test splits used for evaluation.","marker":"[3]"},{"why":"Faster R-CNN with ResNet-50 is the baseline detector whose corrupted versions define the comparison.","marker":"[58]"},{"why":"Provides the uniform bounding-box noise protocol and the MS-COCO evaluation setting the paper follows.","marker":"[12]"},{"why":"OA-MIL is a key comparison method for robust detection with inaccurate boxes.","marker":"[14]"},{"why":"LNCIS is a comparison method for learning with noisy instance labels in detection tasks.","marker":"[13]"},{"why":"Supplies the small-loss criterion and Co-teaching baseline that the paper compares LLS against.","marker":"[9]"},{"why":"Mix-Up is the main augmentation alternative compared and contrasted with patch-level Mix-Paste.","marker":"[27]"},{"why":"CutMix is the patch-replacement augmentation that Mix-Paste distinguishes itself from.","marker":"[55]"}],"fun_headline_variants":["Mix same-label patches to tame noisy X-ray annotations","Blending K patches exponentially cuts label noise","Mix-Paste: robust X-ray detection under 60% noise","Patch mixing beats annotation noise in X-ray screening"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The synthetic noise used in experiments, with random label replacement and uniformly perturbed boxes, is assumed to reproduce how real X-ray annotations get corrupted, and if real mistakes are systematic rather than random, the measured gains may not transfer.","fun_headline_variants_meta":{"raw":{"variants":["Mix same-label patches to tame noisy X-ray annotations","Blending K patches exponentially cuts label noise","Mix-Paste: robust X-ray detection under 60% noise","Patch mixing beats annotation noise in X-ray screening"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000335,"raw_usage":{"total_tokens":1920,"prompt_tokens":1071,"completion_tokens":849,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":687,"completion_tokens_details":{"reasoning_tokens":786}},"tokens_in":687,"tokens_out":849,"duration_ms":9183,"temperature":1.0,"reasoning_tokens":786,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:21:08.851033+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-annotate a subset of OPIXray or PIDray with clean labels, inject structured noise such as confusion concentrated between visually similar categories and boxes consistently inflated in one direction, and check whether Mix-Paste plus LLS keeps its advantage; if the gain collapses, the independence assumption behind $1-P_c^K$ is the weak link.","supporting_citations":[{"cited_title":"Occluded prohibited items detection: An X-ray security inspection benchmark and de-occlusion attention module,","cited_arxiv_id":null,"evidence_quote":"Supplies the OPIXray benchmark and the real noisy-annotation examples that motivate the method."},{"cited_title":"PIDray: A large-scale X-ray benchmark for real-world prohibited item detection,","cited_arxiv_id":null,"evidence_quote":"Supplies the PIDray benchmark with easy, hard, and hidden test splits used for evaluation."},{"cited_title":"Faster R-CNN: Towards real-time object detection with region proposal networks,","cited_arxiv_id":null,"evidence_quote":"Faster R-CNN with ResNet-50 is the baseline detector whose corrupted versions define the comparison."},{"cited_title":"Towards Noise-resistant Object Detection with Noisy Annotations","cited_arxiv_id":"2003.01285","evidence_quote":"Provides the uniform bounding-box noise protocol and the MS-COCO evaluation setting the paper follows."},{"cited_title":"Robust object detection with inaccurate bounding boxes,","cited_arxiv_id":null,"evidence_quote":"OA-MIL is a key comparison method for robust detection with inaccurate boxes."},{"cited_title":"Learning with noisy class labels for instance segmentation,","cited_arxiv_id":null,"evidence_quote":"LNCIS is a comparison method for learning with noisy instance labels in detection tasks."},{"cited_title":"Co-teaching: Robust training of deep neural networks with extremely noisy labels,","cited_arxiv_id":null,"evidence_quote":"Supplies the small-loss criterion and Co-teaching baseline that the paper compares LLS against."},{"cited_title":"CutMix: Reg- ularization strategy to train strong classifiers with localizable features,","cited_arxiv_id":null,"evidence_quote":"CutMix is the patch-replacement augmentation that Mix-Paste distinguishes itself from."}],"review_version":1}