{"id":"f73c32a9-a7f3-41bb-9a7d-57595b70aa3d","arxiv_id":"2505.05936","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"CGTrack, a lightweight one-stream UAV tracker using cascade concatenation with residual squeeze-and-excitation gating and a gated center head, reports state-of-the-art success and precision scores on UAV123, UAV123@10fps, and UAVTrack112.","lead":"CGTrack is a new drone-tracking algorithm that fuses hierarchical image features with cascade gating modules and reports top accuracy on three aerial tracking benchmarks while running at real-time speed. A generalist reader might care because faster, more accurate on-drone tracking is a building block for delivery, inspection, and robotics applications.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline SOTA results are internally inconsistent: Section IV-B reports UAV123@10fps precision 83.8% and UAVTrack112 precision 80.6%, while Fig. 5 reports 86.2% and 82.4% for the same measures; the central claim is not verifiable until this conflict is resolved.","rationale":"The reader's weakest assumption was external baseline comparability, which is a valid concern but not the most load-bearing one here. The more immediate problem is internal: the paper reports conflicting precision numbers for two of the three benchmarks in the main results, and the ablation table does not support the claimed gain. These conflicts directly affect the central SOTA claim because the claimed margins are small enough that 2-6 point discrepancies change whether CGTrack is first or below other trackers. The issue is concrete, falsifiable, and addressable by re-running the evaluation, so the verdict should remain CONDITIONAL rather than REJECT: if the Fig. 5 numbers are confirmed, the paper is largely salvageable with text corrections; if they are not, the SOTA claim fails. I agree with the reader that the paper is not acceptable as stated, but I locate the decisive weakness in the paper's own inconsistent reporting rather than in unstated baseline protocols, which is why agreement is only partial.","tokens_in":12543,"tokens_out":3682,"duration_ms":33085,"concrete_test":"Obtain the final CGTrack-B checkpoint and re-run the official UAV123@10fps and UAVTrack112 evaluation toolkits, recording the exact precision values. If the output matches Fig. 5 (86.2 and 82.4), correct Section IV-B; if it matches Section IV-B (83.8 and 80.6), correct Fig. 5. If it matches neither, audit the evaluation pipeline and checkpoint before any SOTA claim is made.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim rests on the specific SOTA numbers in Fig. 5 and Section IV-B. However, the paper gives two different precision values for two of the three benchmarks. Section IV-B states UAV123@10fps achieves 'highest Precision (83.8%)' while the corresponding precision plot in Fig. 5 labels CGTrack as [86.2]. Likewise, Section IV-B claims a 'Precision score of 80.6%' on UAVTrack112, while Fig. 5 labels CGTrack as [82.4]. Only UAV123 is consistent (88.0% in both). This is not a minor typo in a caption: the claimed SOTA margins over the runner-up on UAV123@10fps are 6.3 points (86.2 vs 79.9) by Fig. 5 but only 3.9 points (83.8 vs 79.9) by the text, and on UAVTrack112 the margin is 1.0 point (82.4 vs 81.4) by Fig. 5 but the text value would actually drop CGTrack below TCTrack++ and SiamAPN (81.4). A related inconsistency appears in the HFC ablation (Table I): the paper says Row 3 gives a 3.34% increase over Row 2, but 86.24-84.12=2.12. Since the paper provides no code, weights, or raw logs, the reader cannot determine which set of numbers is correct. The SOTA conclusion may survive after re-evaluation, but as written the central evidence for it is self-contradictory.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes CGTrack, a family of one-stream UAV trackers built on a lightweight hierarchical LeViT backbone. The two main architectural contributions are a Hierarchical Feature Cascade (HFC) module, which concatenates multi-scale backbone features and applies residual Squeeze-and-Excitation gating, and a Lightweight Gated Center Head (LGCH) whose Efficient Gating (EG) blocks use Hadamard-product gating in an expanded feature space. The authors report state-of-the-art results on UAV123, UAV123@10fps, and UAVTrack112, with the largest variant running at 42.1 fps, and support the design with ablations on fusion strategy, model scale, and head configuration. As written, however, the headline numbers are internally inconsistent: Section IV-B reports different precision values from Fig. 5 on two of the three benchmarks, and one ablation increment in Table I is arithmetically wrong. These issues prevent the central SOTA claim from being verified in the current version.","tokens_in":12844,"tokens_out":5347,"duration_ms":48908,"significance":"If the reported results hold, the paper makes a useful and practical contribution: it shows that a simple concatenation-plus-gating fusion of hierarchical ViT features, together with a lightweight gated center head, can exceed more complex fusion designs on challenging UAV benchmarks while remaining real-time. The ablation structure is sensible, and the idea of reusing hierarchical features through cascade concatenation is well motivated by DenseNet. The paper also evaluates on three standard UAV benchmarks and compares with 13 baselines, which is a reasonable scope. The main weakness is that the central quantitative evidence is not currently trustworthy because of the numerical inconsistencies described below; the significance of the architectural claim can only be assessed after those numbers are reconciled and the evaluation protocol is clarified.","major_comments":[{"comment":"The headline SOTA numbers are internally inconsistent. Section IV-B reports UAV123@10fps Precision 83.8% and UAVTrack112 Precision 80.6%, while the precision plots in Fig. 5 label CGTrack as 86.2% and 82.4% on the same benchmarks, respectively. Only UAV123 (88.0%) is consistent. The difference is not cosmetic: with the text value, CGTrack would trail TCTrack++ and SiamAPN (both 81.4) on UAVTrack112, and the UAV123@10fps margin over SiamTPN would shrink from 6.3 to 3.9 points. Please correct the tables/text/figures to one consistent set of numbers and, if possible, release per-sequence results or raw logs so the reader can verify them.","section":"Section IV-B, Fig. 5"},{"comment":"The claimed gain of the Residual SE gating is misreported. The text states that Row 3 exhibits a 3.34% increase in Precision over Row 2, but 86.24 - 84.12 = 2.12 percentage points. The 1.22% gain from Row 1 to Row 2 is correct (84.12 - 82.90). Please correct the stated increment and re-evaluate the conclusion that cascade gating is the main driver of the HFC improvement.","section":"Section IV-D, Table I"},{"comment":"The final model configuration appears to be selected on the same benchmark used for the headline results. Section IV-D states that all ablations are performed on UAV123@10fps, and Table III selects the EG upsampling ratio 2 by the highest AUC on that benchmark, while Fig. 5 reports CGTrack's UAV123@10fps success and precision on the same dataset. If the selection and the reported result share the same test set, the reported SOTA numbers are optimistically biased. Please state explicitly whether the ablations were done on a held-out validation split and whether the final evaluation is on a disjoint test protocol.","section":"Section IV-D, Table III"},{"comment":"Comparability of the 13 baseline numbers is not established. The manuscript does not state, for each baseline, the training data, template/search resolution, and evaluation protocol used to produce the plotted scores, nor does it say whether all baselines were re-run under a common codebase or quoted from the original papers. Because several UAV trackers are trained on different data (some on UAV123 itself) and CGTrack is trained on GOT-10k/LaSOT/COCO/TrackingNet, the claimed margins (e.g., 88.0 vs 82.3 precision on UAV123) may reflect protocol differences rather than architectural merit. Please provide a table of baseline settings or re-evaluate all methods in a common protocol.","section":"Fig. 5, Section IV-A"},{"comment":"The efficiency claim 'without additional parameters or FLOPs' for the HFC module is contradicted by the paper's own ablation. Table I shows that adding Residual SE (the gating part of HFC) increases parameters from 40.668M to 41.219M and MACs from 4.323 to 4.324G; the Residual SE block itself contains FC layers, as shown in Eqs. (2)-(5). Please restate the claim to refer only to the concatenation step, or provide a corrected analysis.","section":"Section III-C"}],"minor_comments":[{"comment":"The spacing 'UA V' appears in the title, abstract, and main text (e.g., 'UA V Tracking', 'UA V scenarios'); it should be 'UAV'.","section":"Throughout"},{"comment":"The row numbering is duplicated: both 'EG block-3x' and 'EG block-4x' are labeled as row 4. The rows should be numbered 1 through 5.","section":"Table III"},{"comment":"The symbol S in Eq. (5) is used without a definition; presumably it denotes the channel-wise scaling tensor whose entries are s_c from Eq. (4), but this should be stated explicitly.","section":"Eq. (5)"},{"comment":"The heading 'Gating Mechenism' contains a typo; it should be 'Gating Mechanism'.","section":"Section II-C"},{"comment":"The caption writes 'UA V123@10fp' in the second column label; this should be 'UAV123@10fps'.","section":"Fig. 5 caption"},{"comment":"The phrase 'excluding 1k sequences as convention' is ambiguous; please clarify whether this refers to the standard GOT-10k validation split or to some other exclusion.","section":"Section IV-A"}],"recommendation":"major_revision","confidential_remarks":"The core architectural idea is plausible and the ablation design is informative, but the numerical inconsistencies are central to the SOTA claim and must be resolved before publication. I would not accept the paper in its current form, but I see no basis for rejection: the issues appear fixable by correcting the reported numbers, clarifying the evaluation protocol, and ideally releasing code or raw logs. The revision should be checked carefully for consistency between the text, tables, and figures."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. CGTrack is a competent recombination of known ideas into a lightweight one-stream UAV tracker, and the ablations mostly support the individual design choices. But the central SOTA claim is not currently verifiable: Section IV-B reports 83.8% precision on UAV123@10fps and 80.6% on UAVTrack112, while Fig. 5 reports 86.2% and 82.4% for the same measures. The 80.6 value would actually place CGTrack below two baselines in Fig. 5, so this is not a cosmetic typo.\n\nWhat is genuinely new here is the specific configuration: a LeViT one-stream backbone, a cascade concatenation path with residual SE gating, and a gated center head built from StarNet-style Hadamard blocks. The paper credits those sources honestly. The HFC and LGCH ablations (Tables I and III) support the core choices, and the three variants give a useful accuracy/speed spread, from 9.99M params at 61 fps to 41.2M at 42 fps. The benchmark coverage is standard but appropriate for UAV tracking.\n\nThe soft spots are real but not all equal. The text/plot number conflict is load-bearing and has to be fixed before any SOTA sentence can be trusted. The ablation arithmetic is also sloppily described: Row 3 vs Row 2 is 2.12 points, and the claimed 3.34 only works if the comparison is to Row 1, which is not what the sentence says. That is fixable. A broader worry is that the 13 baseline numbers are quoted from their original papers without stating training data, resolution, or evaluation protocol; that matters because CGTrack is trained on a different mixture and some margins, like the 1.0-point lead on UAVTrack112, are thin. No code, weights, or error bars are provided, so independent verification is impossible. The model selection on UAV123@10fps via Table III before claiming SOTA on that same benchmark is a mild circularity, but it is common and not disqualifying.\n\nIf I were handling this, I would send it to peer review with a clear instruction: reconcile the numbers, state baseline protocols, and release code and weights. The architecture idea deserves a controlled comparison against HiT, and the paper is worth a serious referee, but I would not cite the SOTA claims until the inconsistency is resolved.","headline":"The headline results are internally inconsistent (text vs. plots disagree on two benchmarks), so the SOTA claim is not verifiable as written; the architecture is worth a careful look once the numbers are reconciled.","tokens_in":13482,"tokens_out":2811,"would_cite":false,"duration_ms":29945,"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":"CGTrack claims state-of-the-art accuracy on three UAV tracking benchmarks using a cascade gating design that fuses hierarchical features at negligible extra cost.","keywords":["UAV tracking","visual object tracking","hierarchical feature fusion","gating mechanism","lightweight vision transformer","one-stream tracker","feature reuse","real-time tracking"],"falsifier":"A reader could settle it by training CGTrack-B on the four stated training sets and running the same evaluation scripts on UAV123; if CGTrack's precision and success do not reach the reported 88.0% and 67.2%, or if any of the 13 baselines, re-run under the same script, match or exceed those numbers, the state-of-the-art claim fails.","tokens_in":12255,"feed_emoji":"🎯","tokens_out":15165,"duration_ms":133694,"temperature":0.7,"pith_summary":"Unmanned-aerial-vehicle (UAV) tracking must run in real time on limited hardware, which pushes designers toward lightweight networks that often lose the fine detail needed to keep a target through occlusion, low resolution, and sharp viewpoint changes. The paper introduces CGTrack, a one-stream tracker—template and search image are processed jointly in a single network—built on a lightweight hierarchical vision transformer, and argues that two additions close that gap: a Hierarchical Feature Cascade (HFC) module that concatenates multi-scale features and applies residual channel gating, and a Lightweight Gated Center Head (LGCH) that uses Hadamard-product gating to extract target-specific detail from the expanded features. The claim is that this concatenate-then-gate design expands network capacity with almost no added parameters or FLOPs, and the paper reports state-of-the-art numbers on UAV123, UAV123@10fps, and UAVTrack112, with the largest variant running at 42.1 frames per second. If true, the work points to feature reuse and gating as a cheap way to get hierarchical fusion benefits in resource-constrained trackers rather than relying on heavier transformer fusion.","feed_headline":"Cascade gating tracker tops UAV123 at 88% precision","feed_subtitle":"CGTrack beats 13 prior trackers on three UAV benchmarks at 42 fps.","key_machinery":"The load-bearing mechanism is the Hierarchical Feature Cascade (HFC) module: it upsamples each backbone stage's correlation map, concatenates adjacent scales along the channel dimension, and passes the output through a Residual Squeeze-and-Excitation block that rescales channels by a learned gating vector, then repeats the cascade once more. Concatenation is the key operation—it preserves every channel's information where addition would force a lossy sum, and the residual SE applies coarse gating at minimal cost. On top of the HFC runs the Lightweight Gated Center Head (LGCH), the final head that predicts center location and bounding-box size; its Efficient Gating (EG) blocks split an input into a gate branch and a context branch through parallel $1\\times1$ convolutions, pass one branch through ReLU6, and combine them with an element-wise (Hadamard) product, mapping the features into a higher-dimensional nonlinear space while computing in the low-dimensional one. The backbone provides three hierarchical correlation maps from a lightweight LeViT, and the loss is the standard one-stream combination of focal, GIoU, and $\\ell_1$ terms, so the HFC-plus-LGCH pairing is what carries the argument.","core_discovery":"On the paper's own terms, the central claim is that a cascade of feature concatenation and gating, rather than addition-based or transformer-based fusion, lets a lightweight hierarchical ViT compete with heavier trackers in aerial scenarios. The paper reports that CGTrack reaches 88.0% precision and 67.2% success on UAV123, 86.2% precision and 66.1% success on UAV123@10fps, and 82.4% precision and 64.9% success on UAVTrack112, surpassing the 13 published trackers it is compared with in Fig. 5. The design rationale is explicit: concatenation preserves the rich spatial detail of shallow stages alongside semantic depth, Residual Squeeze-and-Excitation blocks reweight the channels, and the gated center head decouples target-oriented coordinates in a higher-dimensional nonlinear feature space. Table I's ablation on UAV123@10fps shows the full cascade-gating configuration 3.34 precision points above addition-based fusion and 2.12 points above concatenation without residual SE, which is the paper's main evidence that the HFC mechanism, not just the backbone, drives the improvement.","pith_inferences":["A plausible generalization is that the same concat-then-gate cascade could serve as a universal fusion block in other one-stream trackers or multi-scale detection heads, and the strongest test would be reproducing the Table I gain outside this architecture.","The attribute results suggest a testable prediction the paper does not fully exploit: the gated cascade should matter most on small-object and occlusion attributes, so a controlled study varying object size could isolate where HFC's contribution comes from.","A reader should note an internal inconsistency: Section IV-B's prose reports 83.8% precision on UAV123@10fps and 80.6% on UAVTrack112, while Fig. 5 plots 86.2% and 82.4%; the plotted values should be treated as the canonical claim.","If the HFC design transfers, the same fusion trick could be applied to other resource-constrained multi-scale tasks like detection or segmentation, where concatenation and gating are usually studied separately."],"forward_implications":["If the reported numbers hold, a lightweight one-stream tracker can reach state-of-the-art aerial accuracy without the transformer-based fusion layers that heavier UAV trackers use, since CGTrack-B's 4.324 GMACs keep it in real-time range.","The HFC module is a drop-in style intervention: swapping additive fusion for concatenation-plus-residual-gating is what Table I credits for a 3.34-point precision gain on UAV123@10fps, so the same swap should transfer to other trackers built on hierarchical backbones.","The three variants, CGTrack-T, -S, and -B, span 61.4 to 42.1 fps with precision from 80.08 to 86.24 (Table II), giving practitioners a direct speed-accuracy trade-off for different drone-hardware budgets.","Because CGTrack trains only on GOT-10k, LaSOT, COCO, and TrackingNet, its evaluation numbers are not in-domain, so the reported SOTA is on unseen aerial test sets rather than benchmarks the model has seen."],"supporting_citations":[{"why":"It supplies DenseNet's dense-connection feature-reuse idea that the HFC module simplifies into a cascade of concatenations.","marker":"[35]"},{"why":"It supplies LeViT, the lightweight hierarchical ViT backbone whose three-scale features CGTrack feeds into the HFC.","marker":"[47]"},{"why":"It supplies HiT, the one-stream hierarchical ViT tracker whose framework and dual-image position encoding CGTrack adapts.","marker":"[32]"},{"why":"It supplies the one-stream tracking formulation (joint feature learning and relation modeling) and the loss weights $\\lambda_G=2$, $\\lambda_l=5$ CGTrack adopts.","marker":"[15]"},{"why":"It supplies StarNet's account of gating/Hadamard-product features, the basis for the EG block's higher-dimensional nonlinear mapping.","marker":"[39]"},{"why":"It supplies SENet's channel-wise Squeeze-and-Excitation gating, which CGTrack recasts as a residual block inside the HFC.","marker":"[44]"},{"why":"It supplies UAV123 and UAV123@10fps, the two benchmarks where the paper's state-of-the-art claims are made.","marker":"[34]"},{"why":"It supplies UAVTrack112, the third benchmark, and the SiamAPN baseline that motivates the anchor-free comparison.","marker":"[48]"},{"why":"It supplies TCTrack, one of the 13 baselines whose reported scores CGTrack must beat on the UAV benchmarks.","marker":"[42]"},{"why":"It supplies Ocean, one of the strongest published baselines on UAV123, used as a direct comparison in Fig. 5.","marker":"[62]"}],"fun_headline_variants":["Cascade gating lifts UAV tracking to 88% precision","Feature reuse + gating: CGTrack beats 13 UAV trackers","CGTrack: lightweight cascade gating tops UAV benchmarks","Gated cascade tracker wins at 42 fps on UAV benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The state-of-the-art conclusion rests on the assumption that the published scores of the 13 compared trackers were produced under the same training data, resolution, and evaluation protocol as CGTrack; if any baseline came from a different protocol, the reported margins would not be a fair measure of the architecture.","fun_headline_variants_meta":{"raw":{"variants":["Cascade gating lifts UAV tracking to 88% precision","Feature reuse + gating: CGTrack beats 13 UAV trackers","CGTrack: lightweight cascade gating tops UAV benchmarks","Gated cascade tracker wins at 42 fps on UAV benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00022,"raw_usage":{"total_tokens":1479,"prompt_tokens":1010,"completion_tokens":469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":626,"completion_tokens_details":{"reasoning_tokens":396}},"tokens_in":626,"tokens_out":469,"duration_ms":4671,"temperature":1.0,"reasoning_tokens":396,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T22:54:09.248597+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A reader could settle it by training CGTrack-B on the four stated training sets and running the same evaluation scripts on UAV123; if CGTrack's precision and success do not reach the reported 88.0% and 67.2%, or if any of the 13 baselines, re-run under the same script, match or exceed those numbers, the state-of-the-art claim fails.","supporting_citations":[{"cited_title":"LeViT: a Vision Transformer in ConvNet’s Clothing for Faster Inference,","cited_arxiv_id":null,"evidence_quote":"It supplies LeViT, the lightweight hierarchical ViT backbone whose three-scale features CGTrack feeds into the HFC."},{"cited_title":"Exploring lightweight hierarchical vision transformers for efficient visual track- ing,","cited_arxiv_id":null,"evidence_quote":"It supplies HiT, the one-stream hierarchical ViT tracker whose framework and dual-image position encoding CGTrack adapts."},{"cited_title":"Joint Feature Learning and Relation Modeling for Tracking: A One-Stream Framework,","cited_arxiv_id":null,"evidence_quote":"It supplies the one-stream tracking formulation (joint feature learning and relation modeling) and the loss weights $\\lambda_G=2$, $\\lambda_l=5$ CGTrack adopts."},{"cited_title":"Rewrite the stars,","cited_arxiv_id":null,"evidence_quote":"It supplies StarNet's account of gating/Hadamard-product features, the basis for the EG block's higher-dimensional nonlinear mapping."},{"cited_title":"A Benchmark and Simulator for UA V Tracking,","cited_arxiv_id":null,"evidence_quote":"It supplies UAV123 and UAV123@10fps, the two benchmarks where the paper's state-of-the-art claims are made."},{"cited_title":"Onboard real-time aerial tracking with efficient siamese anchor proposal network,","cited_arxiv_id":null,"evidence_quote":"It supplies UAVTrack112, the third benchmark, and the SiamAPN baseline that motivates the anchor-free comparison."},{"cited_title":"Tctrack: Temporal contexts for aerial tracking,","cited_arxiv_id":null,"evidence_quote":"It supplies TCTrack, one of the 13 baselines whose reported scores CGTrack must beat on the UAV benchmarks."},{"cited_title":"Ocean: Object-aware Anchor-free Tracking,","cited_arxiv_id":null,"evidence_quote":"It supplies Ocean, one of the strongest published baselines on UAV123, used as a direct comparison in Fig. 5."}],"review_version":1}