{"id":"c4588a1d-7105-402a-af34-5a242e044bbe","arxiv_id":"2411.18082","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"A new large-scale dual-view X-ray dataset and a two-pipeline detection framework improve detection of hard categories by exploiting a second X-ray view.","lead":"This paper introduces a large dual-view X-ray dataset for prohibited item detection, with 353,646 labeled instances across 12 categories, and a detection framework that uses both vertical and side views to improve detection of hard categories. The method reports large gains on tail categories, such as a 24.7 point improvement in umbrella detection, and the dataset is positioned as a new benchmark for security inspection AI.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No controlled single-view ablation: reported dual-view gains may come from added crop/expert refinement, not from the second view.","rationale":"The reader's conditional verdict is appropriate and should not be changed. The paper contributes a very large dual-view dataset and reports consistent gains across seven detectors, but the central claim that the second view is responsible for those gains is confounded by the simultaneous introduction of expert crop refinement. This is not an accusation of error; it is a missing control that could be supplied quickly. The reader's stated weakest assumption concerns geometric alignment (Eq. 1), which is a real transferability risk, but the more immediate correctness risk is that the dual-view contribution is not causally isolated. If the proposed replacement-view control shows no drop in UM AP50, then the paper's main claim would need to be reframed as 'expert crop refinement improves detection' with dual-view proposals as one possible proposal source. Until that control is run, the conditional verdict remains the right call, and the condition should explicitly include the single-view ablation.","tokens_in":15106,"tokens_out":4430,"duration_ms":43379,"concrete_test":"Run AENet under the exact protocol of Table 3 and Table 5, but with the auxiliary view replaced by a second copy of the main view (I2 := I1), keeping the same saliency detector, auxiliary detector, expert models, merging strategy, and confidence threshold. If the UM AP50 still rises from 33.9 to roughly 59.0, the improvement is caused by the crop/expert/refinement pipeline, not by dual-view information, and the paper must add this control before claiming a dual-view benefit. Repeat for all seven base detectors in Table 3 to check whether the reported consistency also holds in the single-view control.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the second X-ray view is what drives the large AP50 gains on hard categories (e.g., UM from 33.9 to 59.0 in Table 3). The experimental design does not isolate this factor. Moving from the single-view RetinaNet row to the '+ours' row changes two things at once: (i) it adds a saliency-based proposal and auxiliary detector operating on I2, and (ii) it adds crop-level expert classifiers trained on ground-truth main-view patches for hard categories (Sec. 4.3). The single-view baseline has neither component. The gain could therefore be produced entirely by the crop/refinement machinery on I1, with I2 only supplying proposals that a saliency map on I1 could also supply. Table 5's 'Saliency' row does not resolve this, because it still uses I2 as the proposal source and lacks the control of running the same pipeline with I1 in place of I2. Furthermore, the geometry in Eq. 1 is under-specified and may not transfer across scanner geometries, but the attribution confound is more fundamental: without a single-view control, the headline claim that the dual-view mechanism significantly enhances detection is not supported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces LDXray, a large paired dual-view X-ray dataset containing 146,997 image pairs and 353,646 annotated instances across 12 prohibited-item categories, and proposes AENet, a framework that augments a standard main-view detector with an auxiliary-view pipeline. The auxiliary pipeline uses saliency detection on the second view to generate proposals, pseudo-labels them via a presumed horizontal-coordinate correspondence with the main view, trains an auxiliary detector, and then crops corresponding main-view regions for refinement by category-specific expert classifiers. Experiments on LDXray report consistent AP50 gains over seven base detectors, with the largest gains on hard categories such as umbrellas (e.g., RetinaNet AP50 from 33.9% to 59.0%). The paper also reports ablations on expert model choice, saliency-based localization, auxiliary detector choice, and confidence thresholds.","tokens_in":15299,"tokens_out":6228,"duration_ms":59618,"significance":"If the central attribution claim is established, the contributions are valuable: LDXray appears to be by far the largest dual-view X-ray detection benchmark, and the reported gains are consistent across seven detector architectures on a held-out test set, with code and data promised for release. The evaluation is not circular: expert models are trained on main-view ground-truth crops, and the auxiliary pipeline's pseudo-labels derive from saliency and label transfer on training data, not from test labels. However, the manuscript does not currently isolate the role of the second view: the single-view baseline differs from the full AENet by several components at once, so the headline claim that the dual-view mechanism drives the improvement is not yet supported by the experiments.","major_comments":[{"comment":"The central claim that the dual-view mechanism drives the large AP50 gains is not supported by the experimental design. Comparing the base detector row with the '+ours' row changes multiple factors simultaneously: the auxiliary view I2 is added, but so is a saliency proposal stage, an auxiliary detector trained with pseudo-labels, a projection of auxiliary boxes onto I1, and crop-level expert classifiers for hard categories. There is no control that runs the same AENet pipeline with I1 in place of I2, and no comparison to a simple two-view fusion baseline or to an existing dual-view method such as DvXray. Table 5's 'Saliency' row does not resolve this confound: it only varies the location-approximation technique inside the auxiliary pipeline and still uses I2 as the proposal source while retaining expert refinement. Please add a single-view control (all AENet components, but with proposals generated from I1 instead of I2) and at least one non-AENet dual-view fusion baseline. If the single-view control reproduces the umbrella gains of 20-25 percentage points, the claim should be reframed as 'crop/expert refinement improves hard-category detection,' not 'the dual-view mechanism improves detection.'","section":"Section 5.1.1, Table 3"},{"comment":"The cross-view correspondence model is under-specified. Equation (1), b*_i(x1,x2)(I1) = λ b*_i(x1,x2)(I2), assumes that a single constant scale factor λ relates horizontal coordinates between the two views across the entire image. This is an exact statement only under strong assumptions about scanner geometry (e.g., parallel cameras with no perspective effects and no depth variation). Real X-ray scanners with perspective projection, sensor tilt, or objects at different depths in the luggage will produce position-dependent λ values. Because the auxiliary pipeline crops main-view regions based on this projection, a violation would misalign the expert-refinement crops and directly reduce the reported gains. Please define the notation x1, x2 explicitly, state the scanner geometry that makes Eq. (1) exact, and provide empirical validation such as the distribution of residuals between projected auxiliary-view box centers and ground-truth main-view box centers on the test set.","section":"Section 4.1, Eq. (1)"},{"comment":"The ablation in Table 5 is described as comparing location-approximation techniques, but the 'Single-view' row appears to be a different system from the 'Conversion' and 'Saliency' rows, and the table caption does not define what each row contains. Moreover, these rows report COCO-style mAP and AP75 while Table 3 and the surrounding text report AP50; the switch in metric is not explained, making the numbers appear internally inconsistent (e.g., AP50 62.7 vs. 65.1 while mAP is 38.2 vs. 39.1). Please state the metric for each table explicitly, define the 'Single-view' and 'Conversion' settings precisely, and make clear that this table does not provide the isolated single-view control needed for the paper's central claim.","section":"Section 5.2.2, Table 5"}],"minor_comments":[{"comment":"The caption of Table 4 lists every expert model as '[24]' (Faster R-CNN); please cite each architecture correctly, e.g., C-RCNN [2], S-RCNN [28], RetinaNet [16], CenterNet [5], RepPoints [36], and ATSS [38].","section":"Section 5.2.1, Table 4"},{"comment":"The notation in Eq. (1) and Eq. (3) is ambiguous: b*_i(x1,x2) is called the 'abscissa' of the bounding box, but the notation looks like a two-dimensional box, and Eq. (3) later uses B*_I2(x1,x2) as if it were a set of boxes. Please define the variables x1, x2 and consistently distinguish scalar coordinates from full bounding boxes.","section":"Section 4.1"},{"comment":"In the provided manuscript, the text inside Figure 6 appears as encoded glyph sequences (e.g., '/uni00000013/uni00000011/...'), making the figure unreadable. Please ensure fonts are embedded correctly and redraw the figure with legible axis labels and legend.","section":"Section 5.2.4, Figure 6"},{"comment":"There are minor spelling inconsistencies in category names (e.g., 'Columnar Orange Liquid' appears with variant spellings in different places); please unify the naming and abbreviations across the text, tables, and figures.","section":"Section 3.1"}],"recommendation":"major_revision","confidential_remarks":"The dataset is a potentially significant resource for the X-ray security inspection community, and the consistent gains across seven detectors are encouraging. The main barrier is the attribution confound: the experiments do not yet demonstrate that the second view, rather than the added crop/expert refinement machinery, is responsible for the reported improvements. This is fixable with additional controlled experiments, so I recommend major revision rather than rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a dataset paper with a real resource, plus a method that reports consistent, large gains on tail categories like umbrellas. The dataset alone justifies reading. But the experimental design does not actually isolate the dual-view contribution, so the headline claim is overstated.\n\nWhat's new: LDXray has 146,997 paired images and 353,646 instances across 12 categories, collected from real security environments with consent and de-identification, annotated by professional inspectors. That is a substantial step up from earlier dual-view benchmarks like DvXray and Dualray. The evaluation uses a held-out test set across seven base detectors, and the gains on hard categories (e.g., umbrella AP50 from 33.9 to 59.0 with RetinaNet) are consistent, which is evidence that something in the pipeline works.\n\nSoft spots: the central attribution is not established. Moving from single-view RetinaNet to '+ours' changes two things at once: it adds the auxiliary-view proposal source (saliency on I2 plus an auxiliary detector) and it adds crop-level expert classifiers trained on ground-truth main-view patches. Table 5's 'Saliency' row still uses I2 as the proposal source; there is no control where the same crop/refinement machinery runs on proposals from I1. The gains could therefore come entirely from the expert refinement on crops, with the second view being incidental. Also missing: comparisons to existing dual-view methods (DvXray, Dualray) or even a simple two-view fusion baseline. The geometric relation in Eq. 1 (a constant horizontal scale factor between the two views) is plausible for a fixed scanner geometry but under-specified, and it is unclear how it transfers to other scanners. Code and dataset are not yet public, which blocks reproduction.\n\nIn proportion, these are fixable issues. The dataset is a contribution regardless, and a proper single-view control plus baseline comparisons would settle the method's actual contribution.\n\nWho this is for: researchers working on X-ray security inspection and multi-view detection. A serious referee should engage — the dataset deserves peer review, and a revision round could clarify the method's real advantage.","headline":"LDXray is a genuinely useful dual-view X-ray benchmark, but AENet's headline claim that the dual-view mechanism drives the gains is not isolated in the experiments.","tokens_in":15871,"tokens_out":1792,"would_cite":true,"duration_ms":16566,"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 second X-ray view, combined with saliency proposals and category-specific expert detectors, substantially improves detection of difficult prohibited-item categories such as umbrellas.","keywords":["dual-view X-ray detection","prohibited item detection","X-ray security inspection","object detection","AENet","LDXray dataset","saliency detection","cross-view correspondence"],"falsifier":"Collect dual-view images from a scanner whose side view is tilted or whose objects sit at different depths, run the trained AENet without adaptation, and measure the umbrella AP50 gain; if the gain disappears because the projected main-view crops are misaligned, the fixed-scale coordinate assumption is the failing component.","tokens_in":14896,"feed_emoji":"🧳","tokens_out":7275,"duration_ms":62896,"temperature":0.7,"pith_summary":"The paper claims that the human practice of checking both a vertical and a side X-ray view can be reproduced by a two-pipeline detection network. The main-view pipeline detects ordinary categories; a saliency-based auxiliary-view pipeline proposes candidate regions in the side view, transfers class labels through a fixed horizontal-coordinate correspondence, and passes cropped main-view regions to expert models trained on hard categories. The claim is supported by a new dual-view dataset, LDXray, with 146,997 image pairs and 353,646 instances across 12 categories, and by consistent gains on seven base detectors. The largest reported gain is for umbrellas, whose AP50 rises from 33.9% to 59.0% when RetinaNet is paired with AENet.","feed_headline":"Second X-ray view lifts hard-item detection by 25 points","feed_subtitle":"A new 353k-instance dual-view dataset and a two-pipeline network match how inspectors use a side view.","key_machinery":"The load-bearing identity is the cross-view coordinate correspondence b*_i(x1,x2)(I1) = lambda b*_i(x1,x2)(I2), together with the class-label correspondence c*_i(I1)=c*_i(I2). This lets an unlabeled auxiliary view be converted into training signal: saliency detection proposes object locations, labels are borrowed from the main view, and the same horizontal projection places those locations back onto the main view to define crops. The auxiliary detector and the category-specific expert models are the two learned components that carry the refinement, because the auxiliary detector localizes candidates and the expert models confirm their category on high-resolution main-view crops.","core_discovery":"The authors set out to show that dual-view X-ray images can be used the way human inspectors use them, and they present AENet as the mechanism that does so. The auxiliary view has no ground-truth annotations, so the framework generates pseudo-labels: saliency detection proposes boxes in I2, Eq. 1 transfers the main view's class labels to those boxes under a single scale factor lambda, and the auxiliary detector learns from these labels. At inference, the auxiliary boxes are merged, projected to the main view, cropped, and refined by expert models; the final output unifies the main-view and refined auxiliary detections. The reported result is that hard categories improve substantially, with umbrella AP50 climbing by 20.1 to 25.1 percentage points across base detectors, while already-easy categories stay at the same level.","pith_inferences":["A direct consequence the paper does not develop is that the same cross-view projection could serve as a self-supervised pretraining signal on raw dual-view images, without any inspector annotations.","If the fixed-scale lambda assumption is violated by other scanner geometries, the reported transfer gains would likely shrink; a per-image or per-depth calibration of the projection would be the natural repair.","The dataset's two views could support a stricter test than the paper reports: measuring whether items invisible in the main view are recovered from the auxiliary view, which would isolate the unique contribution of the second view.","One could test the saliency proposal stage independently by checking whether the projected auxiliary boxes cover the ground-truth main-view boxes on a held-out set; that coverage number would predict how much of the umbrella gain is due to localization versus expert classification."],"forward_implications":["Adding AENet improves AP50 for the umbrella category by 20.1 to 25.1 percentage points across the seven base detectors, while leaving the simple categories' AP50 unchanged.","The gain is architecture-independent: Faster R-CNN, Cascade R-CNN, Sparse R-CNN, RetinaNet, CenterNet, RepPoints, and ATSS all report higher mAP when the auxiliary-view pipeline is added.","Saliency-based location approximation outperforms a direct coordinate-conversion baseline as the auxiliary-view proposal method, with mAP rising from 38.2 to 39.1 on the RetinaNet setup.","The auxiliary pipeline can recover very hard categories; using Cascade R-CNN as the auxiliary detector yields 78.6% AP50 for umbrellas in that branch alone.","LDXray, at 146,997 image pairs and 353,646 instances, is offered as the first large-scale open dual-view detection benchmark for this task."],"supporting_citations":[{"why":"GDXray is the foundational single-view X-ray detection benchmark that Table 1 uses to position LDXray's scale and dual-view novelty.","marker":"[22]"},{"why":"SIXray is the large single-view security-inspection benchmark used as a comparison point in the dataset overview.","marker":"[23]"},{"why":"Faster R-CNN is one of the seven base detectors whose hard-category AP50 the AENet pipeline improves.","marker":"[24]"},{"why":"RetinaNet is the base detector behind the headline umbrella gain, from 33.9% to 59.0% AP50.","marker":"[16]"},{"why":"Cascade R-CNN serves as both a main-view base detector and the best-performing auxiliary-view and expert model in the ablations.","marker":"[2]"},{"why":"The open-source detection toolbox supplies the consistent training and evaluation environment for every base detector.","marker":"[3]"},{"why":"MV-Xray is the earlier multi-view X-ray detection work whose dual-view setting LDXray extends to much larger scale.","marker":"[27]"},{"why":"Dualray is the other dual-view detection benchmark listed in Table 1, used to compare dataset size and availability.","marker":"[34]"},{"why":"The standard annotation quality-control protocol is the procedure the authors follow to double-check inspector annotations.","marker":"[6]"}],"fun_headline_variants":["Dual-view X-ray AI lifts hard item detection by 25 points","Two X-ray views help AI spot prohibited items like human inspectors","Umbrella detection jumps 25 points with dual-view X-ray AI","AI side-view trick: hard items found with 25-point AP gain","Dual-view X-ray network mimics human inspection, boosts hard categories by 25"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline depends on the assumption that a single fixed scale factor maps an object's horizontal position in the auxiliary view to its horizontal position in the main view across the whole image.","fun_headline_variants_meta":{"raw":{"variants":["Dual-view X-ray AI lifts hard item detection by 25 points","Two X-ray views help AI spot prohibited items like human inspectors","Umbrella detection jumps 25 points with dual-view X-ray AI","AI side-view trick: hard items found with 25-point AP gain","Dual-view X-ray network mimics human inspection, boosts hard categories by 25"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000691,"raw_usage":{"total_tokens":3122,"prompt_tokens":932,"completion_tokens":2190,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":2095}},"tokens_in":548,"tokens_out":2190,"duration_ms":15833,"temperature":1.0,"reasoning_tokens":2095,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:31:12.090256+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect dual-view images from a scanner whose side view is tilted or whose objects sit at different depths, run the trained AENet without adaptation, and measure the umbrella AP50 gain; if the gain disappears because the projected main-view crops are misaligned, the fixed-scale coordinate assumption is the failing component.","supporting_citations":[{"cited_title":"Gdxray: The database of x-ray images for nondestructive testing","cited_arxiv_id":null,"evidence_quote":"GDXray is the foundational single-view X-ray detection benchmark that Table 1 uses to position LDXray's scale and dual-view novelty."},{"cited_title":"Sixray: A large-scale security inspection x-ray benchmark for prohibited item discovery in 9 overlapping images","cited_arxiv_id":null,"evidence_quote":"SIXray is the large single-view security-inspection benchmark used as a comparison point in the dataset overview."},{"cited_title":"Faster r-cnn: Towards real-time object detection with region proposal networks","cited_arxiv_id":null,"evidence_quote":"Faster R-CNN is one of the seven base detectors whose hard-category AP50 the AENet pipeline improves."},{"cited_title":"Cascade r-cnn: Delv- ing into high quality object detection","cited_arxiv_id":null,"evidence_quote":"Cascade R-CNN serves as both a main-view base detector and the best-performing auxiliary-view and expert model in the ablations."},{"cited_title":"Multi- view x-ray r-cnn","cited_arxiv_id":null,"evidence_quote":"MV-Xray is the earlier multi-view X-ray detection work whose dual-view setting LDXray extends to much larger scale."},{"cited_title":"Dualray: Dual-view x-ray security inspec- tion benchmark and fusion detection framework","cited_arxiv_id":null,"evidence_quote":"Dualray is the other dual-view detection benchmark listed in Table 1, used to compare dataset size and availability."},{"cited_title":"The pascal visual object classes (voc) challenge","cited_arxiv_id":null,"evidence_quote":"The standard annotation quality-control protocol is the procedure the authors follow to double-check inspector annotations."}],"review_version":1}