{"id":"864b5207-fd84-4fca-ac41-9228b4f2643a","arxiv_id":"1908.03295","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"low","formal_verification":"none","parameter_count":4,"one_line_summary":"PADet combines anchor promotion and feature alignment in a single-stage detector, reaching 40.0 percent mAP on MS COCO test-dev at 28.6 fps.","lead":"This paper describes PADet, a single-shot object detector that uses an anchor promotion module to increase the number of good training boxes and a feature alignment module to keep those promoted boxes aligned with image features. It reports 40.0 percent mAP on MS COCO test-dev at 28.6 frames per second, a competitive speed-accuracy tradeoff for real-time detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"All ablations use VGG-16 at 384×384, while the headline 40.0 mAP claim is ResNet-101 at 512×512; the transfer of APM/FAM gains across backbone and input scale is untested.","rationale":"The reader's weakest assumption is exactly the load-bearing point. All component evidence is at VGG-16/384, while the headline number is at ResNet-101/512. Without a ResNet-101/512 baseline or component ablation, the 40.0 mAP could be attributed to the heavier backbone and larger input rather than to APM and FAM. This is not an internal inconsistency, but it is a genuine evidence gap in the central empirical claim. The proposed ablation would settle whether the modules transfer. Other concerns, such as the small 0.7 mAP margin of option (c) over option (b) in Table 1 or the absence of error bars, are real but less load-bearing: the headline could remain true even if those details changed. The reader's CONDITIONAL verdict is therefore appropriate and should be retained.","tokens_in":11307,"tokens_out":6784,"duration_ms":70691,"concrete_test":"Run the incremental ablation from Table 2 (baseline, +APM_C, +APM_R, +APM, +APM+FAM) with ResNet-101 at 512×512, using the same training schedule, anchor settings, and matching strategy as the headline model. Then check two quantities: (i) the ResNet-101/512 baseline mAP without APM/FAM, and (ii) the incremental gain of FAM over APM at this configuration. If the baseline already approaches 40.0, or if the FAM gain is much smaller than the 3.1 mAP seen at VGG-16/384, the central claim that the modules drive the improvement is not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The component analysis (Tables 1–4) is conducted entirely with the VGG-16 encoder-decoder at 384×384 input. The headline result (abstract and Table 6) is ResNet-101 at 512×512, reaching 40.0 mAP and 28.6 fps. Between these two configurations, both the backbone and input resolution change simultaneously, yet no ResNet-101 baseline or component-wise ablation is reported. The central argument that APM and FAM cause the improvement therefore depends on an unverified transfer assumption: the +2.1 mAP from APM and +3.1 mAP from FAM (Table 2: 31.7→34.8) should remain comparable when the encoder-decoder uses ResNet-101 at 512. This is not a cosmetic gap: APM's anchor scoring and FAM's offset prediction operate at specific feature strides and anchor densities, which change with backbone and input size. The matching statistics motivating the design (Figures 1–2) are computed for VGG-16/384 only. A plausible alternative reading of the headline is that ResNet-101 at 512 would already achieve much of the 40.0 mAP without APM/FAM, with the modules contributing less than the ablation suggests.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes PADet, a single-shot object detector that combines an Anchor Promotion Module (APM) and a Feature Alignment Module (FAM). APM predicts the probability that each anchor is positive and adjusts its location and shape, thereby increasing the number of high-quality positive anchors while suppressing easy negatives. FAM uses the location and shape transformation information from APM to compute disentangled offsets for deformable convolution, aligning features to the promoted anchors. The detector uses an encoder-decoder backbone and is evaluated on MS COCO and PASCAL VOC with both VGG-16 and ResNet-101. The headline result is 40.0% mAP on MS COCO test-dev with ResNet-101 at 512×512 input at 28.6 fps. The ablations, all on VGG-16 at 384×384, show that APM and FAM together improve mAP from 29.6% to 34.8% on the COCO minival set.","tokens_in":11722,"tokens_out":7705,"duration_ms":75210,"significance":"The paper addresses a genuine problem in one-stage detection—foreground/background imbalance—by actively promoting positive anchors rather than only down-weighting easy negatives, which is a useful and timely direction. The FAM design is thoughtful, and the systematic comparison of offset-learning variants in Table 1 and component ablations in Table 2 are strengths. The method is evaluated on the external MS COCO benchmark, and the reported gains do not reduce by construction to the target metric. The release of code and the additional PASCAL VOC comparison add credibility. The main limitation is that all component analysis is confined to one backbone/input configuration, while the headline uses a different configuration, and the speed comparison is not controlled. If the transfer of APM/FAM gains across backbone and input scale can be demonstrated and the ResNet-101 configuration fully specified, the contribution would be solid.","major_comments":[{"comment":"All ablations in Tables 1–4 and the anchor statistics in Figures 1–2 are computed with VGG-16 at 384×384, while the headline 40.0 mAP (abstract and Table 6) is obtained with ResNet-101 at 512×512. No ResNet-101 baseline or component-wise ablation is reported, so the central claim that APM and FAM are responsible for the headline improvement rests on the untested assumption that the +2.1 mAP from APM and +3.1 mAP from FAM (Table 2) transfer when both the backbone and input resolution change. Because anchor stride, anchor density, and matching statistics depend on these settings, please add component ablations for ResNet-101 at 384 and 512 against a ResNet-101 encoder-decoder baseline, or at least verify that the VGG-16 gains persist at 512×512 before attributing the headline numbers to the modules.","section":"Section 4 and Table 6"},{"comment":"The architecture description covers only the reduced VGG-16 backbone. Table 6 reports 'Ours ResNet-101 384×384' and 'Ours ResNet-101 512×512', but the paper never specifies how the encoder-decoder is adapted to ResNet-101: which ResNet stages are used, what strides and channels the decoder levels have, what anchor settings are used, and whether the training schedule differs from the VGG-16 setting. This missing specification prevents reproduction and verification of the central result. Please provide the full ResNet-101 configuration and any implementation differences.","section":"Section 3, Backbone Network"},{"comment":"The speed comparison is not controlled: the authors' FPS is measured on a GTX 1080 Ti with PyTorch 0.4.1, while the comparison values are quoted from other papers and libraries using different hardware and software. The abstract's claim of running at 28.6 fps and being 'faster' than alternatives is therefore not directly supported, and the '7.5× faster than RetinaNet-101-800' statement in Section 1 has the same issue. Several comparison rows also lack FPS values entirely (e.g., SSD ResNet-101, GA-RetinaNet, FSAF). Please either measure the comparison methods in the same environment or explicitly frame the speed numbers as indicative and soften the implied controlled comparison.","section":"Section 4.2, Table 6"},{"comment":"The threshold θ=0.01 is central to the method's design, since it controls the positive/negative balance during training and the filtering of detections at inference, but no sensitivity analysis is reported, and all mAP values come from single runs. Because the main claimed gains are 2–3 mAP, it is important to know the effect of θ on the balance and final mAP, and ideally the variance across seeds. Please add a θ sweep (including θ=0) and, if possible, repeated runs for the main configurations in Tables 2 and 6.","section":"Section 3.1 and Section 4"}],"minor_comments":[{"comment":"The sentence claiming results are obtained 'without multi-scale training and any testing tricks' is inconsistent with Section 3.3, which applies soft-NMS at inference; soft-NMS is generally considered a post-processing/testing technique and should be disclosed in that sentence.","section":"Section 4.2"},{"comment":"The text describes c as 'a scalar' in the disentangled offset formulation, but the module outputs 2 channels for c (an x,y translation) and 2K channels for Δs; please correct the terminology to avoid confusion.","section":"Section 3.2"},{"comment":"The text says adding the feature alignment module 'improves the performance by 5.2% mAP', but the table shows the full model (APM+FA) at 34.8 versus the APM-only model at 31.7, a 3.1-point gain; the 5.2-point figure appears to be the total gain from the 29.6 baseline. Please clarify which comparison is intended.","section":"Table 2 and Section 4.1"},{"comment":"There is a typo: 'RestinaNet-101-800' should be 'RetinaNet-101-800'.","section":"Section 1"},{"comment":"The text refers to 'Figure 5 (b∗)' but the figure panels are labeled (a), (b), and (c); please align the references with the panel labels.","section":"Figure 5 and text after it"},{"comment":"The paper acknowledges that small-object improvements decrease rapidly with IoU and attributes this to only 34.3% of small ground-truth objects having more than 2×2 pixels at stride 8; this limitation is stated clearly, but it would be helpful to also report the number of small objects at each IoU range so the reader can assess the statistics.","section":"Section 4.1, Table 4"}],"recommendation":"major_revision","confidential_remarks":"The paper's core idea is reasonable and the VGG-16 ablations are informative, but the missing ResNet-101 baseline and the uncontrolled speed comparison are the main risks. If the authors can supply the requested ResNet-101 ablations and a full architecture specification, the paper would be a strong candidate for acceptance. I do not see evidence of circularity or misrepresentation beyond the 'no testing tricks' phrasing noted in the minor comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear [Colleague], quick take on arXiv:1908.03295. The paper is a competent one-stage detector contribution. The genuinely new piece is the disentangled offset decomposition in the Feature Alignment Module: a shared scalar offset from location change plus a residual vector from shape change, instead of concatenating the two. Table 1 shows this beats implicit, explicit, and concatenated alternatives, and the ablation story in Table 2 (APM and FAM each help, combined best) is coherent. The authors are honest that FAM is not applied to the last two layers and that they use a lightweight head. The PASCAL VOC results and COCO numbers are plausible. The speed-accuracy tradeoff at 28.6 fps / 40.0 mAP is attractive if it holds up.\n\nThe soft spots are real but not disqualifying. The stress-test note is correct: all component ablations use VGG-16 at 384×384, while the headline is ResNet-101 at 512×512. There is no ResNet-101 baseline without APM/FAM, so the exact contribution of the modules at that scale is unverified. That said, the paper does report ResNet-101 at 384×384 (36.9 mAP), which is an intermediate data point—so the authors are not hiding the backbone switch entirely. Still, a referee should ask for a ResNet-101 baseline and ideally a partial ablation at 512. The theta=0.01 threshold and soft-NMS parameters are tuned on minival; that is standard practice but a sensitivity check on theta would be easy and would strengthen the paper. No error bars or repeated runs—typical for this literature, so I would not weight that heavily. The speed comparison relies on published numbers for other methods and their own machine for theirs; that is also standard, but the claims about '7.5x faster' should be taken as approximate.\n\nCitation pattern is fine: they build on RefineDet, SRN, GA-RPN, DCN, and cite them properly. No self-citation issues. The claim that RefineDet filters out ~10% hard positives is not sourced in the text, but it is a minor point.\n\nWho is this for? Anyone working on single-shot detectors, especially anchor-based refinements and feature alignment. The disentangled offset idea is worth citing. I would give this a serious peer review: the empirical work is systematic, the writing is clear, and the main gap is a missing baseline rather than a broken argument. My recommendation: send it out, and ask for the ResNet-101 ablation and a theta sensitivity experiment. Verdict for me: conditional accept—I believe the approach works, but I want the missing baseline before I'd trust the headline number.","headline":"Solid single-shot detector paper with a genuine module-design insight; the main gap is that the headline ResNet-101 result lacks a no-module baseline.","tokens_in":12086,"tokens_out":1702,"would_cite":true,"duration_ms":16209,"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 single-shot detector closes the gap to two-stage accuracy by promoting positive anchors and aligning features, reaching 40.0% mAP on MS COCO at 28.6 fps.","keywords":["single-shot object detection","class imbalance","anchor promotion","feature alignment","deformable convolution","MS COCO","foreground-background balance"],"falsifier":"Re-run Table 2's ablation (baseline, +APM_C, +APM_R, +APM, +APM+FAM) on ResNet-101 with 512×512 input on COCO minival: if the FAM increment over APM is well below 5.2 mAP, or if the full model falls short of 40.0% on test-dev, the scale-transfer assumption fails. A sharper check is to replace the disentangled offset ($c + \\Delta s$) with the concatenated $F(X_l; X_s)$ at the full scale and compare the gap with the 0.7 mAP gap reported at 384×384.","tokens_in":11103,"feed_emoji":"🎯","tokens_out":6775,"duration_ms":60641,"temperature":0.7,"pith_summary":"This paper argues that one-stage object detectors lose accuracy to two-stage detectors mainly because their anchor boxes are poorly placed: huge numbers of background boxes dominate training while genuine positive boxes are rare and imprecisely located. Rather than only down-weighting easy negatives as focal loss does, the paper proposes actively promoting anchors—predicting which anchors are likely positive and adjusting their location and shape so that more of them become true positives. It then adds a feature alignment module that samples features at the promoted anchors' new positions and scales using disentangled offsets from the promotion branch. Together the two modules lift a VGG-16 detector from 29.6% to 34.8% mAP on COCO minival at 384×384, and a ResNet-101 detector to 40.0% mAP on test-dev at 28.6 fps. If correct, this shows the class-imbalance problem in dense detectors is better solved by creating positives than by suppressing negatives alone.","feed_headline":"Promoted anchors push single-shot detector to 40% mAP","feed_subtitle":"New modules for anchor promotion and feature alignment match two-stage accuracy at 28.6 fps.","key_machinery":"The Anchor Promotion Module (APM) and Feature Alignment Module (FAM). APM predicts a per-anchor positive score and regresses location and shape adjustments, so training the detection head sees many more and better-aligned positive boxes while easy negatives are discarded via a score threshold $\\theta = 0.01$. FAM then applies deformable convolution, with offsets computed as a shared scalar translation $c$ learned from the location-change features plus a residual vector $\\Delta s$ learned from the shape-change features (a \"disentangled\" decomposition), so that the features sampled for classification and regression correspond to the promoted anchors rather than the original grid. The two modules sit on an encoder-decoder backbone with skip connections, attaching to each decoder level.","core_discovery":"The central claim is that a dense single-shot detector can be made substantially more accurate by treating anchor boxes as editable entities: an Anchor Promotion Module (APM) scores each anchor's probability of being positive and regresses offsets to its center and size, turning some negative anchors into positive ones and improving the IoU of existing positives; this lifts the positive-to-negative ratio from roughly 1:625 to 1:10 during training. Because the promoted anchors no longer sit at original feature-grid locations, the paper introduces a Feature Alignment Module (FAM) that pools features using deformable convolution whose offsets are computed disentangled—a shared translation from the location-change branch and a per-kernel residual from the shape-change branch—rather than implicitly or by concatenating the two cues. On MS COCO test-dev the combined system reaches 40.0% mAP with ResNet-101 at 512×512 input and 28.6 fps, and ablation shows FAM contributes a 5.2 mAP gain over APM alone.","pith_inferences":["The \"promote positives\" recipe should transfer to anchor-free detectors: replacing grid points with predicted object-center candidates, then sampling features with the same disentangled offsets, could recover the small-object recall gains without dense anchor tiling.","The disentangled offset decomposition—shared translation plus per-kernel residual—is a general way to inject known geometric transformations into deformable convolution; it could be applied to tracking or instance segmentation where motion or shape priors are available explicitly.","The paper's own observation that gains shrink at IoU > 0.7 suggests that a cascade of promotion-and-alignment stages, trained at progressively higher IoU thresholds, might extend the method toward high-quality detection the way cascade R-CNN does for two-stage detectors.","The ablation indicates the scoring branch (APM_C, +1.4 mAP) carries more weight than the regression branch alone (APM_R, +1.0 mAP), but the biggest jump comes from combining both and then aligning features, implying the two functions are complementary rather than redundant."],"forward_implications":["At 384×384 with VGG-16, adding FAM on top of APM raises mAP from 31.7% to 34.8%, showing that alignment is what makes promoted anchors usable.","On MS COCO test-dev, the ResNet-101 512×512 model reaches 40.0% mAP at 28.6 fps, outperforming RetinaNet-101-800 (37.8 mAP, ~5 fps) with several times the speed.","Anchor promotion nearly quadruples small-object mAP in the ablation setting (12.1% to 18.6%), suggesting the method specifically addresses the recall problem that hurts small objects in dense detectors.","Discarding low-score negatives at inference (scores below $\\theta = 0.01$) means the promoted anchors and their scores are used end-to-end, not just as a training-time sampling trick.","The ratio of positives to negatives during training improves from roughly 1:625 to 1:10, which is the direct mechanism the paper credits for easing the class imbalance."],"supporting_citations":[{"why":"Provides the encoder-decoder backbone used as the baseline 'vanilla DSSD' in the ablation study.","marker":"[37]"},{"why":"RetinaNet with focal loss is the main competing treatment of class imbalance and the speed/accuracy comparison target.","marker":"[7]"},{"why":"RefineDet's anchor refinement module suppresses negatives; the proposed APM is contrasted with it.","marker":"[8]"},{"why":"SSD supplies the anchor design, matching rules, and data augmentation used throughout.","marker":"[22]"},{"why":"Deformable Convolutional Networks provide the sampling mechanism that FAM builds on.","marker":"[14]"},{"why":"Faster R-CNN's RPN is the proposal mechanism that APM is compared against and distinguished from.","marker":"[3]"}],"fun_headline_variants":["Anchor Promotion Module lifts single-shot to 40 mAP","Promoted anchors and aligned features yield 40 mAP","Editable anchors balance positives: 40 mAP at 28.6 fps","Feature alignment with promoted anchors: 40 mAP","Single-shot detector with APM+FAM hits 40 mAP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The component-wise gains are measured only with VGG-16 at 384×384 input; the headline 40.0% mAP uses ResNet-101 at 512×512 with no per-component ablation, so the paper assumes APM and FAM transfer their gains to the heavier backbone and larger input unchanged.","fun_headline_variants_meta":{"raw":{"variants":["Anchor Promotion Module lifts single-shot to 40 mAP","Promoted anchors and aligned features yield 40 mAP","Editable anchors balance positives: 40 mAP at 28.6 fps","Feature alignment with promoted anchors: 40 mAP","Single-shot detector with APM+FAM hits 40 mAP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000249,"raw_usage":{"total_tokens":1529,"prompt_tokens":901,"completion_tokens":628,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":541}},"tokens_in":517,"tokens_out":628,"duration_ms":5940,"temperature":1.0,"reasoning_tokens":541,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:17:50.300071+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run Table 2's ablation (baseline, +APM_C, +APM_R, +APM, +APM+FAM) on ResNet-101 with 512×512 input on COCO minival: if the FAM increment over APM is well below 5.2 mAP, or if the full model falls short of 40.0% on test-dev, the scale-transfer assumption fails. A sharper check is to replace the disentangled offset ($c + \\Delta s$) with the concatenated $F(X_l; X_s)$ at the full scale and compare the gap with the 0.7 mAP gap reported at 384×384.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RefineDet's anchor refinement module suppresses negatives; the proposed APM is contrasted with it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"SSD supplies the anchor design, matching rules, and data augmentation used throughout."},{"cited_title":"Faster R-CNN: towards real-time object detection with region proposal networks","cited_arxiv_id":null,"evidence_quote":"Faster R-CNN's RPN is the proposal mechanism that APM is compared against and distinguished from."}],"review_version":1}