{"id":"6f71e948-9004-4472-9e71-0d8f6dc7d215","arxiv_id":"2411.15772","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Corner2Net detects objects as cascade corners, predicting class-agnostic top-left corners first and instance-specific bottom-right corners in each RoI, and claims state-of-the-art among corner-based detectors.","lead":"A new object detector, Corner2Net, finds objects by first spotting every possible top-left corner, then using that corner to locate the matching bottom-right corner, avoiding the error-prone corner-matching step of earlier methods. It reports top results among corner-based detectors on COCO and runs faster, but it leaves code unpublished and uses a non-standard pedestrian benchmark protocol.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Corner2Net's SOTA claim rests on first-stage top-left/scale quality; Table 5 shows the cascade collapses without the enlarge slack, so the reported margin is not yet robustly established.","rationale":"The reader's weakest assumption identifies exactly the same load-bearing concern: the cascade pipeline assumes the first stage supplies a reliable top-left corner and a scale whose enlarged square RoI contains the true bottom-right corner. The paper's own Table 5 provides strong internal evidence for this concern, since performance collapses when the enlarge slack is removed. I agree with the conditional verdict: the idea is plausible and the framework is coherent, but the headline claim of a 'large margin' in accuracy is not fully supported without quantifying first-stage errors, releasing code/weights, and providing variance estimates. The small AP gap over CornerAffinity makes this fragility decisive rather than cosmetic. No fatal flaw was found, and the concern is not about external consensus; it is about internal error-accumulation risk that the paper itself demonstrates but does not analyze.","tokens_in":13113,"tokens_out":5711,"duration_ms":56251,"concrete_test":"Use the trained model on COCO val: (1) measure top-left recall among the top-128 extracted proposals against ground-truth top-left corners; (2) for each ground-truth object, compute the predicted square RoI with eta=0.25 and record the fraction of ground-truth bottom-right corners that fall outside it; (3) feed oracle RoIs (ground-truth top-left and ground-truth scale, with and without small perturbations) into the second stage and compare the resulting AP with the end-to-end AP. If the oracle AP is much higher, first-stage errors dominate and the cascade bottleneck is confirmed; if not, the second stage is the limiting factor. Additionally, rerun the Table 5 ablation with eta=0.25 for at least five random seeds to quantify variance and verify that the margins exceed run-to-run noise.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that Corner2Net surpasses all existing corner-based detectors by a large margin depends on the cascade pipeline's ability to place the true bottom-right corner inside the square RoI defined by the first-stage top-left corner and the regressed scale, expanded by eta=0.25 (Eq. 4). This is a serial composition: if the top-left corner is missed, or if the scale estimate is too small by more than the fixed slack, the second stage cannot recover the object because the bottom-right corner is searched only inside that RoI. The paper's own ablation (Table 5) demonstrates the load-bearing nature of this assumption: setting eta=0 drops AP from 45.7 to 32.5 on COCO val, and eta=0.10 still loses 0.7 AP relative to eta=0.25. This implies that a substantial fraction of instances have first-stage errors that would place the bottom-right corner outside the un-enlarged RoI. Yet the paper provides no analysis of top-left recall at the top-128 truncation, no distribution of scale regression errors, and no error bars over training runs. Since the reported accuracy advantage over CornerAffinity is only about +0.5 AP in the single-scale comparison, the 'large margin' claim is not separated from this fragile first-stage dependence. The design is coherent and the ablation is honest, but the strong SOTA claim is conditional on an unquantified error-accumulation bottleneck.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Corner2Net, a two-stage cascade corner-based object detector. In the first stage, class-agnostic top-left corners are detected on FPN heatmaps along with a square RoI defined by a regressed object scale; in the second stage, the instance-specific bottom-right corner is localized within the enlarged RoI and the object category is predicted from RoI features. The method is evaluated on MS-COCO, CityPersons, and UCAS-AOD. The authors report 46.8/47.8 AP on COCO test-dev with ResNeXt-101-DCN under single/multi-scale testing and claim to surpass all existing corner-based detectors in accuracy and speed.","tokens_in":13387,"tokens_out":6393,"duration_ms":51686,"significance":"The cascade-corner formulation is a genuinely different way to sidestep heuristic corner matching, and the paper demonstrates that it can be trained with standard backbones in far fewer epochs than Hourglass-based corner detectors. The ablation on the enlarge factor (Table 5) is honest and informative, and the high-AP80 results are a useful data point. The speed advantage (8.0 FPS versus 3.7-3.9 FPS for Hourglass-based corner detectors) is meaningful. However, the central 'large margin' claim is overstated: the single-scale AP gap over the best prior corner-based detector (CornerAffinity) is only 0.5 AP, and the cascade's dependence on first-stage top-left recall and scale accuracy is not quantified, leaving the robustness of the method under-specified.","major_comments":[{"comment":"The claim that Corner2Net 'exceeds all existing corner-based detectors by a large margin' is not supported by the data in Table 1 for single-scale testing: Corner2Net achieves 46.8 AP on COCO test-dev versus 46.3 AP for CornerAffinity, a 0.5 AP difference, under a different backbone (ResNeXt-101-DCN versus Hourglass-104) and much shorter training (24 versus 320 epochs). Even the multi-scale result (47.8 versus 46.3) is a 1.5 AP gap rather than a large margin, and no error bars are reported for any configuration. Please either provide controlled comparisons (same backbone and schedule) and repeated-run variance, or temper the claim to 'comparable or slightly better accuracy with substantially higher speed and training efficiency.'","section":"Section 1, Table 1"},{"comment":"The cascade pipeline's accuracy is critically dependent on the first-stage top-left corner and scale estimate, as shown by the eta ablation: setting eta=0 drops AP from 45.7 to 32.5 on COCO val2017, and eta=0.10 still loses 0.7 AP relative to eta=0.25. The paper does not report the recall of the first-stage top-left detector at the top-128 truncation, the distribution of scale regression errors, or the fraction of instances whose true bottom-right corner falls outside the un-enlarged RoI. Without these diagnostics, the reader cannot separate first-stage error accumulation from the method's localization quality; given the small accuracy margin over prior corner detectors, this gap undermines the robustness of the state-of-the-art claim. Please add these analyses and discuss how often the second stage can recover from first-stage misses.","section":"Section 4.6, Table 5; Section 4.2"},{"comment":"The CityPersons evaluation protocol is not sufficiently described. The sentence 'We filter the vanilla dataset, and merge the main annotations of pedestrians and riders' does not specify which images or annotations are removed, whether the standard 'reasonable' subset and evaluation settings are used, or whether AP is computed on visible or full-body boxes. Because Table 3 compares against methods whose CityPersons numbers may have been obtained under different protocols, the reported improvements of 36.2% and 18.0% over CornerNet on AP50 are not verifiable. Please specify the exact filtering, annotation merging, evaluation metric, and test split, and preferably use the standard CityPersons evaluation protocol.","section":"Section 4.1, Table 3"}],"minor_comments":[{"comment":"In the GIoU loss formula, the arguments of LGIoU appear to be duplicated: it reads LGIoU[(zj, zj), (zf_j, zf_j)] instead of comparing the predicted scale zj with the ground truth zf_j. Please correct the notation.","section":"Eq. (3)"},{"comment":"The abstract highlights 'remarkable AP80/AP90 of 44.6%/22.4%', but Table 2 shows that the AP90 of Corner2Net (22.4%) is lower than that of CornerNet (23.4%). The main text acknowledges this; the abstract should be adjusted to avoid overstating high-IoU quality.","section":"Abstract, Section 4.3, Table 2"},{"comment":"Please clarify whether the top-128 proposal keypoints are selected per FPN level or globally across all levels, and how multi-level top-left predictions are merged before RoI extraction.","section":"Section 4.2"},{"comment":"Two different methods named 'CenterNet' are listed (reference [30] as center-based and reference [5] as parallel corner-based). The text and table should disambiguate these (e.g., 'CenterNet (Objects as Points)' and 'CenterNet (Keypoint Triplets)') to avoid confusion.","section":"Table 1"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within the journal's scope and the core idea is competent. My main concern is that the headline claims are stronger than the evidence: the accuracy margin is small and the cascade's first-stage dependence is unquantified. I would support a major revision that adds diagnostics (top-left recall, scale-error distribution, recall-within-RoI), clarifies the CityPersons protocol, and revises the 'large margin' phrasing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The one thing to know: this is not a recycled corner detector. Corner2Net's cascade pipeline—class-agnostic top-left corner first, then an instance-specific bottom-right corner searched inside the resulting RoI—is genuinely new relative to CornerNet, CenterNet, CentripetalNet, and CornerAffinity, and it eliminates the fragile corner-matching post-process. That's a real conceptual step, and the paper does it cleanly. The decoupling of localization from classification via RoI features is sensible, and the design lets them swap in common backbones like ResNet and Swin-Transformer, which prior parallel corner detectors couldn't do without the Hourglass. The ablation study is honest: Table 5 shows the enlarge factor eta matters a lot (AP drops from 45.7 to 32.5 at eta=0), and Table 4 shows insensitivity to top-k, which is useful information.\n\nWhere the paper oversells: the 'large margin' claim. On COCO test-dev, Corner2Net gets 46.8 AP with ResNeXt-101-DCN, while CornerAffinity gets 46.3 with Hourglass-104. That's 0.5 AP, and the backbones differ, so the comparison is not controlled. The speed comparison has the same confound—of course a ResNeXt model with FPN runs faster than Hourglass-104. Also, the CityPersons filtering protocol is described in one sentence and not reproducible, there are no error bars or multiple runs reported, and no code is released. The stress-test concern about serial error accumulation is real: if the first stage misses a top-left corner or underestimates scale beyond the eta slack, the second stage can't recover. That's a structural property of the method, not a flaw in the math, but the paper never quantifies top-left recall or scale-error distribution, so we don't know how often this bottleneck bites.\n\nOverall, the method is plausible and the design is well-motivated. The paper deserves a serious referee and could become a useful baseline for corner-based detection. The needed revision is mostly about evaluation transparency: report variance, document CityPersons filtering, release code, and soften the 'large margin' language to something the evidence actually supports. I'd take a look at a revised version, and I'd probably cite it as a baseline in future work.","headline":"A genuine new cascade-corner idea with honest ablations, but the 'large margin' SOTA claim is overstated once you notice the backbones differ and the gap to CornerAffinity is ~0.5 AP.","tokens_in":13949,"tokens_out":1472,"would_cite":true,"duration_ms":15053,"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":"Corner2Net models an object as two cascade corners — a class-agnostic top-left corner found first, then an instance-specific bottom-right corner located inside the region it defines — and reports 47.8% AP on the COCO test-dev, ahead of…","keywords":["object detection","corner-based detection","cascade corners","keypoint heatmaps","RoI feature classification","COCO","corner matching","two-stage detector"],"falsifier":"The paper itself provides the sharpest test: setting the enlarge factor to 0 drops COCO AP from 45.7 to 32.5 on val2017, showing how much the result depends on the enlarged RoI containing the bottom-right corner. A more direct check would be to count, on a labeled set, how often the true bottom-right corner lies outside the predicted RoI and to measure top-left corner recall independently; if either is often wrong, objects are lost before the second stage begins.","tokens_in":12908,"feed_emoji":"🎯","tokens_out":6823,"duration_ms":57531,"temperature":0.7,"pith_summary":"Corner2Net is a proposal to rehabilitate corner-based object detection by changing the order in which corners are found. Instead of predicting top-left and bottom-right corners independently and then matching them heuristically, the detector first finds all class-agnostic top-left corners, uses each to define a square search region scaled by the object's estimated size, and then locates that instance's bottom-right corner inside that region. Object category is assigned separately from rich region-of-interest (RoI) features, so the two corners no longer need to carry class semantics. On the COCO test-dev set the paper reports 47.8% average precision under multi-scale testing, above all previously published corner-based detectors, while running about 2.1 times faster than the CornerNet baseline it builds on. The significance, if the results hold, is that the corner-based paradigm can use standard backbones and avoid fragile matching without giving up accuracy.","feed_headline":"Corner2Net's cascade corners hit 47.8% AP on COCO test-dev","feed_subtitle":"No corner-matching heuristics: class-agnostic corners plus RoI features give the speed and accuracy leap.","key_machinery":"The load-bearing mechanism is the cascade corner association via an enlarged square RoI: a top-left corner and a regressed object scale define a proposal box, and enlarging that scale by $\\eta=0.25$ creates the region inside which the second stage must find that instance's bottom-right corner. This RoI is what links one corner to the other, so the bottom-right corner is no longer searched over the whole image but only inside a region anchored by its partner. Supporting components are the class-agnostic focal loss with distance-aware penalty, bottom-right corner pooling plus deconvolution for the instance-level heatmap, GIoU loss for the scale regression, Smooth L1 losses for the two offsets, and a RoI-feature classification head.","core_discovery":"The central claim is that the two corners of a bounding box should be treated as a cascade, not as a pair of parallel predictions. The first stage outputs a single-channel heatmap of class-agnostic top-left corners on FPN levels, together with position offsets and a per-instance scale; from that it decodes a proposal box with square side length equal to the predicted scale, enlarged by a factor of 0.25 to form an RoI that should contain the true bottom-right corner. The second stage applies RoIAlign and bottom-right corner pooling to produce a 28 by 28 instance-specific heatmap, picks the peak to locate the bottom-right corner, and refines it with an offset. Classification uses the un-enlarged RoI features through a lightweight head, with the final score combining localization confidence and class confidence. The paper claims this design makes corner matching unnecessary, decouples localization from classification, and yields 47.8% AP on COCO test-dev with a ResNeXt-101-DCN backbone, 46.0% with Swin-T, and large gains over CornerNet on crowded and aerial benchmarks.","pith_inferences":["A testable extension of the cascade idea is to replace the single square scale with separate width and height regressions; the square RoI's dependence on one scale is the main internal bottleneck identified by the enlarge-factor ablation, and anisotropic RoIs could recover extreme-aspect-ratio objects.","The class-agnostic first stage could potentially be reused across datasets or even across tasks such as oriented boxes or panoptic segmentation, since only the second-stage head would need retraining.","If the first stage misses a top-left corner, the object is unrecoverable; one repair the paper does not explore is a second pass or a top-k set of top-left candidates per scale, which could be evaluated directly by measuring top-left corner recall."],"forward_implications":["Corner-based detectors no longer need a corner-matching post-process or an instance embedding; a detected box is just a top-left peak plus the bottom-right peak found inside its RoI.","Because the first-stage corners are class-agnostic, the detector can run on a single-channel heatmap and connect to standard FPN backbones, so training takes 24 epochs instead of the 200 to 320 epochs required by hourglass-based corner detectors.","Decoupling classification from the corner heatmaps should improve robustness when similar-looking objects co-occur, which is exactly where the paper reports its largest gains on COCO, CityPersons, and UCAS-AOD.","The reported AP80 of 44.6% and AP90 of 22.4% suggest the cascade produces tightly fitting boxes, consistent with the idea that corners are a precise box representation."],"supporting_citations":[{"why":"CornerNet: the baseline paradigm the paper replaces, and the source of corner pooling, Gaussian heatmap targets, and the main accuracy and speed comparisons.","marker":"[10]"},{"why":"CenterNet: the keypoint-triplet corner-based baseline whose matching approach Corner2Net avoids and which it outperforms.","marker":"[5]"},{"why":"CornerAffinity: the strongest prior corner-matching baseline, surpassed by Corner2Net with far fewer training epochs.","marker":"[26]"},{"why":"FCOS: supplies the FPN multi-level feature assignment used by Corner2Net and provides a same-backbone center-based comparison.","marker":"[24]"},{"why":"Stacked Hourglass: the backbone all previous corner detectors depend on, which Corner2Net replaces with standard backbones.","marker":"[17]"},{"why":"Mask R-CNN: provides RoIAlign, used to extract the RoI features for the bottom-right corner and classification.","marker":"[8]"},{"why":"GIoU loss: the objective used for the regressed instance scale in the first stage.","marker":"[22]"},{"why":"Focal loss: supplies the loss form used for the class-agnostic top-left corner heatmap.","marker":"[12]"},{"why":"COCO: the benchmark dataset for the main results and ablations.","marker":"[11]"}],"fun_headline_variants":["Cascade corners: 47.8% AP without corner matching","Corner2Net: two-step corners beat matching heuristics","Corner2Net decouples localization and classification for 47.8% AP","Class-agnostic corners + RoI features: 47.8% AP on COCO","Corner2Net: cascade pipeline for class-agnostic corners, 47.8% AP"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole pipeline depends on the first stage finding every top-left corner and on the estimated object scale, enlarged by 25%, producing a square region that contains the true bottom-right corner; if either fails, the second stage has no way to recover the object.","fun_headline_variants_meta":{"raw":{"variants":["Cascade corners: 47.8% AP without corner matching","Corner2Net: two-step corners beat matching heuristics","Corner2Net decouples localization and classification for 47.8% AP","Class-agnostic corners + RoI features: 47.8% AP on COCO","Corner2Net: cascade pipeline for class-agnostic corners, 47.8% AP"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000683,"raw_usage":{"total_tokens":3126,"prompt_tokens":997,"completion_tokens":2129,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":613,"completion_tokens_details":{"reasoning_tokens":2022}},"tokens_in":613,"tokens_out":2129,"duration_ms":13955,"temperature":1.0,"reasoning_tokens":2022,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:54:05.449366+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"The paper itself provides the sharpest test: setting the enlarge factor to 0 drops COCO AP from 45.7 to 32.5 on val2017, showing how much the result depends on the enlarged RoI containing the bottom-right corner. A more direct check would be to count, on a labeled set, how often the true bottom-right corner lies outside the predicted RoI and to measure top-left corner recall independently; if either is often wrong, objects are lost before the second stage begins.","supporting_citations":[{"cited_title":"Law and J","cited_arxiv_id":null,"evidence_quote":"CornerNet: the baseline paradigm the paper replaces, and the source of corner pooling, Gaussian heatmap targets, and the main accuracy and speed comparisons."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CenterNet: the keypoint-triplet corner-based baseline whose matching approach Corner2Net avoids and which it outperforms."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"CornerAffinity: the strongest prior corner-matching baseline, surpassed by Corner2Net with far fewer training epochs."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"FCOS: supplies the FPN multi-level feature assignment used by Corner2Net and provides a same-backbone center-based comparison."},{"cited_title":"Newell, K","cited_arxiv_id":null,"evidence_quote":"Stacked Hourglass: the backbone all previous corner detectors depend on, which Corner2Net replaces with standard backbones."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Mask R-CNN: provides RoIAlign, used to extract the RoI features for the bottom-right corner and classification."},{"cited_title":"Rezatofighi, N","cited_arxiv_id":null,"evidence_quote":"GIoU loss: the objective used for the regressed instance scale in the first stage."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Focal loss: supplies the loss form used for the class-agnostic top-left corner heatmap."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"COCO: the benchmark dataset for the main results and ablations."}],"review_version":1}