{"id":"285debdf-7f3c-43f1-a907-4cec55393d23","arxiv_id":"2412.10040","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"RemDet is a real-time UAV object detector whose GatedFFN, ChannelC2f, and CED modules reduce information loss, achieving 40.0 mAP on VisDrone and 110 FPS on a 4090.","lead":"This paper presents RemDet, a real-time object detector for drone images that combines a multiplication-based feed-forward block, a widened C2f neck, and a context-enhanced downsample module. It reports state-of-the-art accuracy on VisDrone and UAVDT while running at high frame rates on a single RTX 4090.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 110 FPS likely applies to single-image inference, not to the 40.0 mAP o+ca configuration; the paper never reports end-to-end latency of the cluster-aware crop pipeline, so the SOTA-accuracy-plus-realtime claim is unsubstantiated.","rationale":"The paper's strongest empirical evidence is Table 1, but the key row conflates two evaluation regimes. The single-image result (RemDet-X 640 o, 29.9 mAP, 8.9 ms) is credible and useful. The 40.0 mAP is obtained with o+ca, which is a multi-pass test-time pipeline in the cited literature. Without a latency breakdown, the claim '40.0 mAP at 9 ms' is not established. The reader's concern about validation-set tuning is real but secondary; the latency/protocol mismatch is more directly load-bearing because it affects the headline quantitative claim. I agree with the reader's conditional verdict: the paper is a plausible incremental contribution, but the comparison protocol and latency measurement must be clarified before the SOTA claim is accepted.","tokens_in":16709,"tokens_out":5099,"duration_ms":52907,"concrete_test":"Reproduce RemDet-X 1024 o+ca exactly: instrument the full pipeline (cluster proposal, cropping, per-crop detector forward passes, merging) and measure wall-clock latency and total FLOPs on one 4090. If total latency exceeds 9 ms by more than a small margin, or if total FLOPs exceeds 182 G, then Table 1 must be corrected to report separate latencies for o and o+ca; this would settle whether the 40.0 mAP is real-time.","verdict_should_be":"UNCHANGED","load_bearing_attack":"In Table 1, RemDet-X 1024 o+ca reports 40.0 mAP, 9.0 ms latency, and 182 GFLOPs. However, 'o+ca' means cluster-aware crops are used during evaluation, a protocol that in prior work (YOLC, UFPMP-Det) involves detecting clusters and running multiple network passes. YOLC's o+ca latency is 441 ms. For RemDet the paper reports only a single latency (9 ms) and a single FLOPs (182 G), without describing the crop-generation cost or how many forward passes occur. If 9 ms is the single-image backbone latency, then the 40.0 mAP configuration is slower than claimed and the 110 FPS headline does not apply to it; if 9 ms is end-to-end, then the o+ca protocol must be radically simpler than YOLC's, yet no description or ablation of this component is given. Either way, the central claim that one configuration is both SOTA (40.0) and real-time (9 ms) is not supported by the reported numbers. This is compounded by the paper's stated rejection of 'heavily handcrafted designs' while relying on an undescribed Cluster-Aware Crops method for the headline result.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"RemDet proposes a real-time UAV object detector built around three components: GatedFFN (a gated multiplication block that reparameterizes two convolutions), ChannelC2f (a wider C2f variant for the neck), and Context Enhanced Downsample (CED, a ViT-inspired patch-merge downsampling module). The paper motivates these designs through an 'information loss' principle derived from information bottleneck theory, and claims state-of-the-art results on VisDrone with a 3.4% improvement over prior methods while reaching 110 FPS on a single RTX 4090. Experiments are reported on VisDrone, UAVDT, and MSCOCO, together with ablations and combinations of the RemDet backbone with other detectors.","tokens_in":16993,"tokens_out":3413,"duration_ms":36277,"significance":"If the reported results hold, the paper makes a useful practical contribution: it shows that a relatively simple, hardware-friendly architecture can match or exceed more complex coarse-to-fine UAV detectors, and the cross-dataset experiments on MSCOCO and UAVDT provide evidence of generalization. The paper also ships code and fairly extensive ablations, which is a strength. However, the headline claim of simultaneous state-of-the-art accuracy and real-time speed rests on an evaluation protocol whose overhead is not described, and the theoretical 'information loss' argument is not rigorous. These issues are fixable but currently make the central claim partly unsubstantiated.","major_comments":[{"comment":"The headline configuration RemDet-X 1024 o+ca reports 9.0 ms latency and 182 GFLOPs, but the paper never specifies whether these figures include the cluster-aware crop (ca) pipeline. In prior o+ca methods (YOLC, UFPMP-Det) the crop generation and multiple network passes are part of the reported latency (441 ms and 152 ms respectively). If the 9.0 ms is single-image latency, then the 40.0 mAP result is not real-time in the same end-to-end sense, and the 110 FPS claim does not apply to the SOTA configuration; if it is end-to-end, the ca procedure must be radically cheaper than in prior work, yet no description, pseudocode, or ablation of the crop pipeline is given. Either way the central claim that one configuration achieves both SOTA accuracy and real-time speed is not supported by the reported numbers.","section":"Table 1 and 'Comparison Results on VisDrone'"},{"comment":"The polynomial argument for multiplication is not a rigorous proof that multiplication gives 'higher dimensionality' or better representations. Counting (d+1)d/2 distinct monomials treats the network as computing symbolic polynomials, but in a learned linear layer the weights are not independent symbols, and the output of Eq. (4) is still a single scalar per output channel. The claim that wT2x is 'discarded but implicitly included' is not formalized, and the conclusion that multiplication is 'more cost-effective than feed-forward networks' is not supported by any latency or FLOP comparison of the two modules at matched accuracy; Figure 4(b) reports only mAP versus expansion, not cost-effectiveness. The empirical results may stand, but the theoretical story needs either a precise statement or a re-framing as intuition.","section":"Section 3.2, Eq. (3)-(4)"},{"comment":"The final architecture choices, including the stage block ratio (3:3:6:3), the GatedFFN expansion factor 3, and the CED channel expansion ratio, are selected by ablations on the VisDrone validation set, and the SOTA numbers in Table 1 are reported on the same validation set. This makes the 3.4% improvement over prior work vulnerable to validation-set overfitting; there is no held-out test set or repeated evaluation protocol. The COCO and UAVDT results provide some cross-dataset evidence, but they do not resolve the VisDrone-specific tuning issue. Please either report results on the VisDrone test set (if available), use a validation split distinct from the tuning split, or at least quantify the sensitivity of the headline gain to the tuned hyperparameters.","section":"Ablation of Overall Design; Figure 7; Appendix F"}],"minor_comments":[{"comment":"The paper consistently types 'UA V' instead of 'UAV' (Abstract, Section 1, Figure 1 caption); this should be corrected.","section":"Throughout"},{"comment":"The label 'Marco Design' appears to be a typo for 'Macro Design'; please fix.","section":"Figure 3"},{"comment":"The footnote defines o, ca, and aug but does not explain the bare 'ca' entry for CDMNet; please clarify whether 'ca' implies 'cluster-aware' evaluation without the original validation set, and how that differs from 'o+ca'.","section":"Table 1"},{"comment":"The statement that Multiplication 'discards the wT2x part' seems to contradict the main-text claim that wT2x is implicitly included in the polynomial output; please reconcile this wording.","section":"Appendix B"},{"comment":"RemDet-L is marked with a dagger for cluster-aware crops, but no latency or FLOPs are given for that configuration; since the UAVDT comparison table reports only AP, please state whether efficiency was measured and, if not, note its absence explicitly.","section":"Table 2"}],"recommendation":"major_revision","confidential_remarks":"The submission fits the scope of a computer vision conference, and the code release plus broad comparisons are commendable. However, the evaluation-protocol ambiguity for the o+ca row and the validation-set tuning for the headline numbers are serious enough that I would not accept without major revision. The novelty is mostly in the engineering combination of known components (GLU, patch merge, inverted bottlenecks) wrapped in an information-loss narrative, so the theoretical weakness in Section 3.2 matters for the paper's contribution. I would advise the editor to require a clear statement of what the reported 9 ms latency includes and to ask for either a server-based evaluation or a separate validation split used once for final reporting."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The base detector is a credible incremental contribution. Replacing the C2f bottleneck with a GLU-style gated FFN, widening the neck channels, and adding patch-merge downsampling are simple, hardware-friendly changes. The ablations and cross-dataset numbers support the base-model gains: RemDet-Tiny beats YOLOv8-N by 2.7 points on VisDrone and 2.2 on COCO without any cropping tricks. That part I believe.\n\nThe problem is the headline. The 40.0 mAP row in Table 1 uses cluster-aware crops ('o+ca'), but the paper never describes how that method is applied to RemDet, how many forward passes it needs, or what the crop-generation cost is. That same row reports 9 ms latency and 182 GFLOPs. Prior cluster-crop detectors, YOLC and UFPMP-Det, report 441 ms and 152 ms end-to-end. Either RemDet's 9 ms excludes the crop pipeline, which makes the 'SOTA at real-time speed' claim false, or RemDet's o+ca is radically cheaper than anything in the literature, which needs a description. As written, that row cannot carry the 3.4% improvement claim. The abstract's '110 FPS' also appears to refer to the plain 640 input, not to the 40.0 mAP configuration; the paper doesn't say so.\n\nThe information-loss theory in Section 3.2 is more motivational than load-bearing. The mutual-information equations don't connect rigorously to the module designs, and the stage ratios and expansion factors were tuned on the same validation set used for the final numbers. These concerns are real but secondary. The component ablations and the COCO generalization give me reasonable confidence in the base architecture. Code is released, which helps.\n\nWho is this for? Someone building a real-time UAV detector should read the GatedFFN and CED sections and try to reproduce the base numbers. The o+ca row should be treated as unverified until the protocol is documented.\n\nI'd send it to peer review. The paper deserves referee time, but it needs a substantive revision: report the true end-to-end latency for the o+ca configuration, or drop the o+ca result from the headline and lead with the base-model comparisons. Reviewers should ask for an o+ca method description and a per-component latency breakdown.","headline":"The base detector is a credible incremental contribution, but the headline 40 mAP @ 9 ms o+ca row conflates single-image latency with the cluster-crop pipeline and needs clarification before the SOTA claim is taken at face value.","tokens_in":17546,"tokens_out":3926,"would_cite":false,"duration_ms":39224,"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":"RemDet, a one-stage anchor-free detector built around the principle of reducing information loss, claims state-of-the-art accuracy on VisDrone while running at 110 FPS on a single GPU.","keywords":["UAV object detection","small object detection","real-time detection","information bottleneck","gated feed-forward network","reparameterization","VisDrone","downsampling"],"falsifier":"Run RemDet-X at 1024 on the original VisDrone validation set without cluster-aware crops, or run UFPMP-Det with the same crop settings; if the 3.4-point margin disappears, the headline state-of-the-art claim does not hold. In the same setup, compare GatedFFN against ConvFFN at a fixed accuracy budget; if the multiplication advantage reverses at expansion factors above those tested, the cost-effectiveness claim is empirically bounded.","tokens_in":16479,"feed_emoji":"🚁","tokens_out":4779,"duration_ms":46754,"temperature":0.7,"pith_summary":"RemDet is a one-stage anchor-free detector that aims to show that real-time UAV object detection can be improved by treating information loss as the main design criterion, rather than by adding handcrafted coarse-to-fine pipelines. It claims that widening channels and using multiplication-based gating give high-dimensional representations that preserve small-object information at low latency. On VisDrone, the largest model reaches 40.0 mAP at 9 ms per image, improving on the previous best result by more than 3.4 points while running at 110 FPS.","feed_headline":"Multiplication-based detector reaches 40 mAP on VisDrone at 110 FPS","feed_subtitle":"RemDet shows that design guided by information loss beats heavier UAV detectors while staying real-time.","key_machinery":"The load-bearing object is GatedFFN, a block that computes two pointwise projections of the input, applies GELU, multiplies them element-wise, then fuses with a 1x1 convolution and reparameterized depthwise convolutions. The paper's analytic point is that multiplication of two projections produces a polynomial with O($d^{2}$) distinct terms at no extra parameter cost compared with summing the same projections, so gating gives a higher-dimensional representation than a standard feed-forward network. The C2f-style ChannelC2f and CED downsample modules are support structures that extend the same information-preservation principle to neck width and resolution reduction.","core_discovery":"The paper's central claim is that information loss, not model capacity or handcrafted region proposal, is the main bottleneck for small, dense object detection in UAV images; consequently, architectures should be designed to preserve information across layers. Three modules implement this: ChannelC2f widens the neck channels, CED replaces simple stride-2 convolutions with a combined depthwise convolution and patch-merge downsampling, and GatedFFN replaces the standard feed-forward block with a reparameterized gated multiplication branch. The paper argues analytically and empirically that element-wise multiplication of two projected feature vectors yields an implicit polynomial expansion with more distinct terms than a comparable MLP, making it a cheaper way to reach high-dimensional representations. On VisDrone, RemDet-Tiny improves over the YOLOv8 baseline by 2.7 mAP at comparable speed, and RemDet-X with cluster-aware crops sets 40.0 mAP at 9 ms latency.","pith_inferences":["The information-loss explanation is underdetermined by the experiments; widening channels and multiplication also change parameter count and training dynamics, so the causal role of 'information loss' specifically would need controlled trials.","The paper's own caveat that large-object AP lags behind suggests the approach is tuned to small-object distributions; on datasets with large objects the gains may shrink or reverse.","A natural test is to ablate GatedFFN against a ConvFFN with equal FLOPs and latency across expansion factors 3, 6, and 9 on a non-UAV dataset to see whether the multiplication advantage is dataset-specific."],"forward_implications":["RemDet-Tiny can serve as a drop-in real-time detector for embedded UAV use, beating YOLO-family detectors of similar size on VisDrone.","The information-loss guideline transfers to other detection backbones: replacing backbones in Faster R-CNN, RetinaNet, and DyHead improves their VisDrone AP.","GatedFFN's multiplication principle suggests that gated multiplicative interactions are a lower-latency route to high-dimensional feature representation than wider MLPs.","The combination with cluster-aware crops shows that handcrafted coarse-to-fine pipelines can be replaced or reduced by a simpler architecture that keeps high speed."],"supporting_citations":[{"why":"Supplies the VisDrone dataset used for all main experiments and the state-of-the-art comparison.","marker":"(Zhu et al. 2021)"},{"why":"Provides the YOLOv8 baseline and C2f structure that RemDet builds on and compares against in ablations.","marker":"(Jocher, Chaurasia, and Qiu 2023)"},{"why":"The information bottleneck principle that motivates the information-loss design guideline.","marker":"(Tishby and Zaslavsky 2015)"},{"why":"The GLU gating idea that becomes the multiplication-based GatedFFN module.","marker":"(Dauphin et al. 2017)"},{"why":"UFPMP-Det, the previous state-of-the-art on VisDrone that RemDet-X claims to surpass by 3.4 mAP.","marker":"(Huang, Chen, and Huang 2022)"},{"why":"YOLC, the source of the cluster-aware crops method used in the 40.0 mAP comparison.","marker":"(Liu et al. 2024)"}],"fun_headline_variants":["Info-loss-aware design lifts UAV detection to 40 mAP at 110 FPS","RemDet: rethinking UAV detection via information loss, hits 40 mAP real-time","Multiplication-based GatedFFN plus info-loss design: UAV detector hits 40 mAP","RemDet: information-loss-guided design hits 40 mAP on VisDrone at 110 FPS"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The reported state-of-the-art comparison assumes that using cluster-aware crops for RemDet-X while most baselines are evaluated on the original validation set is a fair protocol, and that tuning stage ratios and expansion factors on the VisDrone validation set does not inflate the reported mAP.","fun_headline_variants_meta":{"raw":{"variants":["Info-loss-aware design lifts UAV detection to 40 mAP at 110 FPS","RemDet: rethinking UAV detection via information loss, hits 40 mAP real-time","Multiplication-based GatedFFN plus info-loss design: UAV detector hits 40 mAP","RemDet: information-loss-guided design hits 40 mAP on VisDrone at 110 FPS"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000761,"raw_usage":{"total_tokens":3427,"prompt_tokens":1041,"completion_tokens":2386,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":657,"completion_tokens_details":{"reasoning_tokens":2289}},"tokens_in":657,"tokens_out":2386,"duration_ms":19474,"temperature":1.0,"reasoning_tokens":2289,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T16:24:44.920961+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run RemDet-X at 1024 on the original VisDrone validation set without cluster-aware crops, or run UFPMP-Det with the same crop settings; if the 3.4-point margin disappears, the headline state-of-the-art claim does not hold. In the same setup, compare GatedFFN against ConvFFN at a fixed accuracy budget; if the multiplication advantage reverses at expansion factors above those tested, the cost-effectiveness claim is empirically bounded.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the VisDrone dataset used for all main experiments and the state-of-the-art comparison."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the YOLOv8 baseline and C2f structure that RemDet builds on and compares against in ablations."},{"cited_title":"N.; Fan, A.; Auli, M.; and Grangier, D","cited_arxiv_id":null,"evidence_quote":"The GLU gating idea that becomes the multiplication-based GatedFFN module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"UFPMP-Det, the previous state-of-the-art on VisDrone that RemDet-X claims to surpass by 3.4 mAP."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"YOLC, the source of the cluster-aware crops method used in the 40.0 mAP comparison."}],"review_version":1}