{"id":"86d745a3-078e-486c-ae89-98863527279b","arxiv_id":"2412.00626","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MambaNUT uses a Mamba backbone with an adaptive curriculum learning schedule to achieve efficient state-of-the-art nighttime UAV tracking.","lead":"The paper introduces MambaNUT, a nighttime UAV tracking system built on a state space model (Mamba) backbone that runs at 75 FPS with only 4.1 million parameters. It claims state-of-the-art accuracy on three nighttime tracking benchmarks by adding an adaptive curriculum that shifts training from daytime to nighttime data.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Equation (3) in §III-C makes nighttime sampling weight 1/(θe), which decreases with epoch; the ACL curriculum as written reduces nighttime data over training, opposite to the paper's stated day-to-night schedule.","rationale":"The reader's weakest assumption—that the static-image nighttime datasets are not clearly converted into tracking pairs—is a genuine reproducibility gap, and I partially agree with it. However, the more directly falsifiable problem is internal to the paper: Eq. (3) contradicts its own prose. The sampling scheduler is one of the two pillars of the central ACL claim, and as written it would reduce nighttime sampling over time, not increase it. This cannot be resolved by standard practice or by consulting benchmark conventions; it requires either corrected equation or released code. I also noted smaller reporting inconsistencies (75 vs. 72 FPS in Table I; 42 GMac/99M vs. 29.4 GMac/92.9M for DCPT in the text), but those are not load-bearing. The verdict remains CONDITIONAL: the proposed architecture may still work if the equation is a typographical error, but the paper must correct the schedule, release the sampler, or rerun the ablation with the stated curriculum before the ACL contribution can be accepted.","tokens_in":12556,"tokens_out":8392,"duration_ms":81971,"concrete_test":"Obtain the released training code at github.com/wuyou3474/MambaNUT (or ask authors for the sampler) and log the nighttime sampling weight at epochs 1, 150, and 300. If the code uses 1/(θe), retrain the baseline+SS configuration from Table II with the increasing schedule described in the text (e.g., e/θ) and compare NAT2024-1 precision/Norm.Prec/Succ; if gains vanish, the ACL claim is unsupported. If the code uses e/θ, the paper's Eq. (3) must be corrected and the schedule validated before the claim is accepted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The core methodological contribution is the sampling scheduler in Eq. (3). The surrounding text says nighttime weights are obtained by 'dividing by a constant and multiplying by the training epochs,' so the fraction of nighttime data should start small and increase. As typeset, however, the weight is w_d = 1/(θe) for nighttime datasets and 1 for daytime datasets, with θ=150. Since e goes from 1 to 300, the nighttime weight falls from 0.0067 to 0.000022 relative to a constant daytime weight of 1, so the sampler would monotonically starve the model of nighttime data rather than moving from easy (daytime) to hard (nighttime) samples. If the formula is implemented literally, the Table II gains attributed to the sampling scheduler cannot be explained by the described curriculum. If it is a typo for e/θ, then the published equation does not describe the actual method, and the ACL mechanism is not reproducible from the paper. Either way, the central claim that ACL improves nighttime tracking is not supported by the text as it stands.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"MambaNUT is a single-stream Vision Mamba tracker for nighttime UAV tracking. The paper proposes an Adaptive Curriculum Learning (ACL) scheme with two schedulers: a sampling scheduler that reweights daytime and nighttime training data over epochs, and an Adaptive Data Weighted (ADW) loss that weights samples by training-set size and per-instance IoU. The method is evaluated on NAT2024-1, NAT2021, and UAVDark135 against 16 baselines, with ablations, attribute-level analysis, and a real-world test on a Jetson Orin NX. The authors report state-of-the-art precision and success on these benchmarks with 1.1 GMac FLOPs, 4.1 million parameters, and roughly 72-75 FPS.","tokens_in":12783,"tokens_out":9540,"duration_ms":94167,"significance":"If the reported results hold, MambaNUT is a meaningful step toward efficient nighttime UAV tracking: it is the first Mamba-based tracker in this setting, it demonstrates a favorable accuracy-efficiency trade-off, and the ablation study suggests that the proposed curriculum contributes to the gains. The paper's strengths include broad comparisons against recent trackers, explicit efficiency reporting, attribute-level evaluation, a real-world embedded demonstration, and a clearly described one-stream architecture. However, the central ACL mechanism has a formula/prose contradiction, the ADW loss term appears too small in scale to explain the measured improvements, and the construction of nighttime tracking training pairs is unspecified; these issues must be resolved before the claims are reproducible.","major_comments":[{"comment":"The sampling scheduler as printed is not the easy-to-hard curriculum described in the text. Eq. (3) gives nighttime datasets a weight of 1/(θ e), with θ = 150 and e the current epoch; since e runs from 1 to 300, the nighttime weight decreases monotonically from 0.0067 to 0.000022 relative to the daytime weight of 1. This would progressively starve the model of nighttime data, which is the opposite of the stated schedule. The surrounding prose says the weights are obtained 'dividing by a constant and multiplying by the training epochs,' which would instead give e/θ and an increasing nighttime weight. If the implementation uses e/θ, the equation must be corrected; if it uses the printed formula, the Table II gains attributed to the sampling scheduler cannot be explained by the described curriculum. Please provide the exact implemented schedule and reconcile Eq. (3) with the text.","section":"§III-C, Eq. (3)"},{"comment":"The training procedure is not reproducible because the paper never specifies how template-search pairs are constructed from BDD100K-Night, SHIFT-Night, and ExDark. These datasets are not standard tracking datasets: ExDark is a still-image dataset, and while BDD100K and SHIFT contain video, the paper does not state whether nighttime ground-truth boxes are used, how positive pairs are sampled (e.g., same-video temporal proximity versus random static pairs), what frame intervals are allowed, or what augmentation is applied. This is load-bearing because the method is a tracker and the ACL is claimed to improve nighttime tracking; if the pairs are formed from independent still images, the model may not learn temporal correspondence, and the reported transfer to UAV video benchmarks would be unsurprising only if the pairing protocol is meaningful. Please describe the exact pair-sampling protocol, including source of target boxes and any filtering.","section":"§IV-A"},{"comment":"The claimed contribution of the ADW loss scheduler is difficult to reconcile with the stated weight γ = 0.00001. For typical IoU values U_i in [0,1], the terms in Eq. (4) are of order 1-10 (or larger for very low IoU), so γ L_ADW contributes at most about 10^-4 to 10^-3 to the total loss when L_cls, λ_iou L_iou, and λ_L1 L_L1 are of order 1-10. Such a negligible term is unlikely to account for the 1.5-3.0 point improvements attributed to the loss scheduler in Tables II and III. Please report the actual scaling used in the implementation, provide gradient norms or loss-scale ablations, or justify why γ = 0.00001 is not effectively a no-op.","section":"§IV-A, Eq. (8)"}],"minor_comments":[{"comment":"The efficiency numbers are inconsistent: §IV-C says DCPT requires 42 GMacs and 99 million parameters, but Table I reports 29.4 GMacs and 92.9 million; the abstract and Fig. 1 say 75 FPS, §IV-C says 'over 75 fps,' and Table I lists 72 FPS. Please reconcile these values.","section":"§IV-C and Table I"},{"comment":"The notation in Eq. (4) is ambiguous: the expression appears as ω(1-Ui)i, and it should be written as ω_i^{(1-U_i)} to distinguish the exponent from the sample index.","section":"Eq. (4)"},{"comment":"There are several typos that should be corrected: 'whih' in §III-A, 'Visison' in §III-B, 'a Adaptive' in §II, and 'logic ratio' in §III-C (likely 'log ratio').","section":"§III-A, §III-B, §III-C"},{"comment":"Fig. 5 uses the symbol 'MambaNUT*' without defining it in the caption; the text in §IV-D defines it, but the figure should be self-contained.","section":"§IV-E and Fig. 5"},{"comment":"The sentence 'the search region and template sizes set to 256 × 256 and 128 × 128' should be rephrased to state explicitly that the search image is 256 × 256 and the template is 128 × 128, matching the notation in §III-B.","section":"§IV-A"},{"comment":"The ablations are reported without variance across runs; given the stochastic nature of deep tracking training, reporting mean and standard deviation over multiple seeds would make the improvements more convincing.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"To the editor: I see no evidence of circular evaluation or test-set fitting; the benchmarks are held out and the ACL schedule depends only on epoch number and training-set statistics. The main risks are internal consistency (Eq. (3) and the γ scaling in Eq. (8)) and reproducibility (nighttime pair construction, code release). These issues appear fixable within the manuscript's scope, so I recommend major revision rather than rejection. Please ask the authors to double-check the DCPT efficiency numbers and to state the exact schedules used in the released models."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nTwo things to know. The paper's headline result — efficient state-of-the-art nighttime UAV tracking with a Mamba backbone — is plausible, and the benchmark numbers are genuinely strong. The second thing is a real problem: Eq. (3) in §III-C writes the nighttime sampling weight as 1/(θe), which decreases from about 0.0067 to 0.000022 over the 300 epochs, while the surrounding prose says the weight should start small and gradually increase. The text explicitly describes 'dividing by a constant and multiplying by the training epochs,' which would give e/θ, the opposite of the printed formula. So as typeset, the sampling scheduler starves the model of nighttime data instead of feeding it more as training progresses. If the equation is literal, the reported gains from the scheduler cannot be explained; if it's a typo, the published math doesn't describe the actual method. Either way, the central mechanism is not reproducible from the paper as written.\n\nThat said, the paper does real work. It's the first to apply a pure Mamba backbone to nighttime UAV tracking, and the efficiency claims check out: 1.1 GMac, 4.1M parameters, 75 FPS, plus a real-world deployment on a Jetson Orin NX running over 30 FPS. The ablation in Table II shows the sampling scheduler and loss scheduler each add roughly 1.5–2 points on NAT2024-1, which is consistent with the narrative if the equation is corrected. The comparison against 16 trackers is standard, and the precision gains over DCPT are modest but consistent across the three benchmarks.\n\nThe soft spots are real but mostly minor. The training-data construction is underspecified: BDD100K-Night, SHIFT-Night, and ExDark are still-image datasets, and the paper never says how template-search pairs are drawn from them. That's a reproducibility gap, but not fatal since COCO has been used similarly in other trackers. The baseline comparison may be tilted if the published trackers didn't receive the same extra nighttime data; the paper doesn't discuss this. No code release, no error bars — both worth asking for but not deal-breakers.\n\nThis paper deserves a serious referee. The core idea is coherent, the efficiency evidence is solid, and the main flaw (the equation) is fixable in revision. I'd send it out, with instructions to the authors to correct Eq. (3), specify the pair-sampling procedure, and discuss the fairness of the baseline comparison. My guess is that after those fixes it could be a solid acceptance.","headline":"Useful engineering contribution with a credible efficiency story, but the ACL sampling equation as printed does the opposite of what the text claims — the central mechanism needs a correction before the results can be fully trusted.","tokens_in":13268,"tokens_out":3672,"would_cite":false,"duration_ms":31906,"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 pure Mamba tracker with adaptive day-night curriculum learning achieves state-of-the-art results on three nighttime UAV tracking benchmarks at 75 FPS with 4.1 million parameters.","keywords":["nighttime UAV tracking","Mamba","state space model","adaptive curriculum learning","single object tracking","low-light tracking","efficient tracking","day-night domain imbalance"],"falsifier":"Reproduce MambaNUT from Section IV-A using only the information given: form template-search pairs from BDD100K-Night, SHIFT-Night, and ExDark, train for 300 epochs with the ACL schedule, and evaluate on NAT2024-1. If the unstated pair-sampling step cannot be implemented unambiguously, or if removing ACL reduces precision by substantially less than the reported 3.7 points, the central claim that ACL plus Mamba delivers the state of the art would be falsified.","tokens_in":12401,"feed_emoji":"🌙","tokens_out":8491,"duration_ms":71972,"temperature":0.7,"pith_summary":"The paper argues that nighttime UAV tracking does not need image enhancement or domain-adaptation discriminators: a single-stream tracker built on a pure Mamba backbone, trained with an adaptive curriculum that moves from daytime to nighttime data, can outperform specialized nighttime trackers while using far fewer parameters and FLOPs. On three nighttime benchmarks—NAT2024-1, NAT2021, and UAVDark135—the proposed MambaNUT sets new state-of-the-art precision and success scores on most metrics, running at roughly 75 FPS with 4.1 million parameters and 1.1 GMac FLOPs. The authors further show that the adaptive curriculum learning (ACL) strategy is responsible for a 3.7-point precision gain on NAT2024-1 over the same model without it. If true, this points to a more efficient end-to-end design for low-light aerial tracking that is deployable on embedded hardware.","feed_headline":"Mamba tracker tops three nighttime UAV benchmarks at 75 FPS","feed_subtitle":"A single-stream state-space model with an adaptive day-night curriculum beats ViT trackers at a fraction of the compute.","key_machinery":"The central object is the bidirectional Vision Mamba encoder, a selective state-space model (S6) that runs forward and backward scans over tokenized template and search patches, replacing the self-attention of ViT-based trackers and keeping cost linear in sequence length. The second mechanism is Adaptive Curriculum Learning (ACL), composed of two schedulers: the sampling scheduler sets the sampling weight of each nighttime dataset to $1/(\\theta e)$ at epoch $e$ with $\\theta=150$, so the night share grows from small to balanced; the loss scheduler applies the ADW loss $L_{\\text{ADW}} = -\\frac{1}{n}\\sum_{i} \\omega_i^{(1-U_i)} \\log(U_i) - U_i(1-U_i)$, with $\\omega_i = \\log(N_{\\max}/N_j)+0.5$ based on dataset size and $U_i$ the instance IoU. These two schedulers work together to make the model first learn easy daytime features and then concentrate on hard nighttime cases.","core_discovery":"MambaNUT is a one-stream nighttime UAV tracker in which a bidirectional Vision Mamba encoder processes template and search patches jointly, and a lightweight center-based head regresses the target box; the backbone has linear complexity in sequence length, giving 1.1 GMac FLOPs and 4.1 million parameters. To handle the scarcity of nighttime tracking data, the authors add Adaptive Curriculum Learning: a sampling scheduler gradually increases the proportion of nighttime samples over 300 training epochs, and an Adaptive Data Weighted (ADW) loss up-weights small datasets and low-IoU instances. On NAT2024-1 the full model reaches 83.3% precision, 76.9% normalized precision, and 63.6% success, beating 16 prior trackers; on NAT2021 and UAVDark135 it leads in precision and normalized precision, with success rates within 0.2% of the best competitor, DCPT. The authors conclude that a pure Mamba backbone plus curriculum-based day-night training yields the best accuracy-efficiency trade-off among current nighttime UAV trackers.","pith_inferences":["The ACL sampling schedule is a general recipe for any imbalanced day/night or domain-shifted tracking setup; the idea of ramping the minority domain's sampling weight from low to balanced over training does not depend on Mamba and could be grafted onto ViT or CNN trackers.","Because the paper never specifies how template-search pairs are drawn from the static nighttime image datasets, an independent reproduction must first solve that step, and how much of the reported gain survives depends on that unstated choice.","The ADW loss's IoU-weighted modulation could plausibly transfer to other box-regression tasks such as daytime tracking or long-tailed object detection, though the paper only tests it in the night-UAV setting.","The one-stream Mamba design's long-sequence modeling may be a natural fit for long-term tracking where template drift accumulates over thousands of frames; NAT2024-1 includes 40 long-term sequences, but the paper does not separately analyze drift behavior."],"forward_implications":["Nighttime UAV tracking can be done end-to-end without a separate low-light enhancer or a domain-adaptation discriminator, simplifying the deployment pipeline.","Mamba-based backbones with linear complexity are sufficient for accurate night tracking, lowering the compute bar: 1.1 GMac FLOPs, 4.1 million parameters, and over 75 FPS on a GPU, with more than 30 FPS on an onboard Jetson Orin NX.","The two ACL schedulers produce consistent gains over plain fine-tuning on NAT2024-1, improving precision, normalized precision, and success by more than three points each when both are used.","A single lightweight tracker can be competitive across three different nighttime benchmarks, indicating the approach generalizes beyond a single evaluation set."],"supporting_citations":[{"why":"Supplies the Mamba selective state-space model that forms the paper's backbone and linear-complexity core.","marker":"[20]"},{"why":"Provides Mamba-R-Small, the specific Vision Mamba backbone variant used in MambaNUT.","marker":"[57]"},{"why":"Defines the one-stream joint feature-learning and tracking-head architecture that MambaNUT adopts.","marker":"[54]"},{"why":"DCPT is the main nighttime SOTA competitor whose performance and computational cost MambaNUT is designed to beat.","marker":"[19]"},{"why":"Introduces the NAT2024-1 long-term benchmark and the TDA-Track baseline evaluated against.","marker":"[14]"},{"why":"Introduces the NAT2021 nighttime aerial tracking benchmark and unsupervised domain adaptation baselines.","marker":"[13]"},{"why":"Provides the UAVDark135 all-day tracking benchmark used as the third evaluation set.","marker":"[45]"},{"why":"Foundational curriculum learning work that motivates the adaptive sampling scheduler in ACL.","marker":"[29]"},{"why":"Source of the dynamically weighted balanced loss that the ADW loss scheduler adapts.","marker":"[44]"}],"fun_headline_variants":["MambaNUT: state-space tracking beats ViT at low light and low cost","Adaptive curriculum and Mamba backbone boost nighttime UAV tracking","MambaNUT achieves SOTA nighttime UAV tracking with linear complexity","MambaNUT tracks nighttime UAVs at 75 FPS with pure Mamba","Nighttime UAV tracking: MambaNUT's curriculum learns day to night"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The training procedure mixes three static nighttime image datasets into template-search pairs, but the paper never explains how those pairs are sampled, so the entire reported gain depends on an unspecified data-construction step that must transfer to real UAV videos.","fun_headline_variants_meta":{"raw":{"variants":["MambaNUT: state-space tracking beats ViT at low light and low cost","Adaptive curriculum and Mamba backbone boost nighttime UAV tracking","MambaNUT achieves SOTA nighttime UAV tracking with linear complexity","MambaNUT tracks nighttime UAVs at 75 FPS with pure Mamba","Nighttime UAV tracking: MambaNUT's curriculum learns day to night"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000822,"raw_usage":{"total_tokens":3634,"prompt_tokens":1019,"completion_tokens":2615,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":635,"completion_tokens_details":{"reasoning_tokens":2516}},"tokens_in":635,"tokens_out":2615,"duration_ms":19143,"temperature":1.0,"reasoning_tokens":2516,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T05:09:27.448325+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Reproduce MambaNUT from Section IV-A using only the information given: form template-search pairs from BDD100K-Night, SHIFT-Night, and ExDark, train for 300 epochs with the ACL schedule, and evaluate on NAT2024-1. If the unstated pair-sampling step cannot be implemented unambiguously, or if removing ACL reduces precision by substantially less than the reported 3.7 points, the central claim that ACL plus Mamba delivers the state of the art would be falsified.","supporting_citations":[{"cited_title":"Joint feature learning and relation modeling for tracking: A one-stream framework,","cited_arxiv_id":null,"evidence_quote":"Defines the one-stream joint feature-learning and tracking-head architecture that MambaNUT adopts."},{"cited_title":"Dcpt: Darkness clue-prompted tracking in nighttime uavs,","cited_arxiv_id":null,"evidence_quote":"DCPT is the main nighttime SOTA competitor whose performance and computational cost MambaNUT is designed to beat."},{"cited_title":"Prompt-Driven Temporal Domain Adaptation for Nighttime UAV Tracking","cited_arxiv_id":"2409.18533","evidence_quote":"Introduces the NAT2024-1 long-term benchmark and the TDA-Track baseline evaluated against."},{"cited_title":"Unsupervised domain adaptation for nighttime aerial tracking,","cited_arxiv_id":null,"evidence_quote":"Introduces the NAT2021 nighttime aerial tracking benchmark and unsupervised domain adaptation baselines."},{"cited_title":"All-day object tracking for unmanned aerial vehicle,","cited_arxiv_id":null,"evidence_quote":"Provides the UAVDark135 all-day tracking benchmark used as the third evaluation set."},{"cited_title":"Curriculum learning,","cited_arxiv_id":null,"evidence_quote":"Foundational curriculum learning work that motivates the adaptive sampling scheduler in ACL."},{"cited_title":"Dynamically weighted bal- anced loss: class imbalanced learning and confidence calibration of deep neural networks,","cited_arxiv_id":null,"evidence_quote":"Source of the dynamically weighted balanced loss that the ADW loss scheduler adapts."}],"review_version":1}