{"id":"f6bf1ca6-cc56-4192-89f2-f5f0195cc640","arxiv_id":"2411.15761","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"MambaTrack is a mamba-based tracker for night UAV tracking that combines low-light enhancement with language prompts, claiming faster and more memory-efficient state-of-the-art performance on five benchmarks.","lead":"MambaTrack is a new drone tracking system for night videos that brightens dark frames and uses a text description of the target to improve tracking. It reports the best accuracy on five night drone benchmarks while running faster and using less memory than transformer-based competitors, though the comparison is complicated by extra text hints at test time.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SOTA comparison is not apples-to-apples: MambaTrack uses per-video language prompts that baselines were not given, and the ablation shows the language branch is exactly what lifts it above CiteTracker.","rationale":"The reader's weakest assumption correctly identifies the language-prompt asymmetry as the main threat to the SOTA claim. The ablation evidence makes this concern concrete: on UAVDark135, MambaTrack without the cross-modal language branch (Baseline + MLLE) is 54.5 AUC, below CiteTracker's reported 55.3, and only reaches 57.2 after adding CMM with language prompts. This means the claimed advantage over the strongest VL baseline is entirely attributable to information that the baseline was not given. The same pattern appears on NAT2021, where the language branch contributes 1.8 AUC. Because the paper adopts published results for all baselines rather than re-running them with the new prompts, the comparison is not apples-to-apples. The efficiency numbers (2.8x faster, 50.2% memory reduction) and component ablations are useful and plausible, but they do not rescue the headline accuracy comparison. The conditional verdict remains appropriate: the paper should be accepted only if the SOTA claim is re-evaluated with baselines given the same language supervision, or if the claim is revised to reflect the prompt-conditioned setting.","tokens_in":10167,"tokens_out":4498,"duration_ms":42581,"concrete_test":"Obtain the released 518 prompts and the public CiteTracker/JointNLT code, then run them on UAVDark135 and NAT2021 with the prompts inserted in the same way MambaTrack uses them (same template/search region, same language embeddings, same evaluation protocol and hardware). If CiteTracker's UAVDark135 AUC reaches or exceeds 57.2, or its NAT2021 AUC reaches or exceeds 53.2, the SOTA claim is an artifact of extra language supervision. Also report MambaTrack without CMM on all five benchmarks to see whether the visual-only pipeline beats the original baselines without language.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central SOTA claim rests on Table I and Figs. 3–4 comparisons against published numbers. The authors state in §III-A that they manually annotated 518 language prompts because the night UAV datasets lack language annotations, and in §III-B that they 'primarily adopt tracking results of methods from original datasets.' Thus MambaTrack is the only method receiving test-time natural-language input on these benchmarks. This is not a minor protocol detail: Table I shows on UAVDark135 that Baseline+MLLE (visual branch with enhancer, no language) scores 54.5 AUC, below CiteTracker's reported 55.3, while adding CMM with language prompts gives 57.2. On NAT2021, the language branch adds 1.8 AUC. Therefore the reported margin over SOTA VL trackers comes from the language prompts, not from the mamba architecture or the dual enhancement itself. Since CiteTracker and JointNLT are themselves VL trackers, they would plausibly benefit from the same prompts. The paper provides no result with these baselines re-run under the prompt-conditioned protocol, so the headline 'best AUC and mACC' is not yet established.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes MambaTrack, a mamba-based tracker for night UAV tracking that combines two enhancement mechanisms: a mamba-based low-light enhancer (MLLE) built on Retinex theory, and a cross-modal mamba (CMM) network that fuses visual features with language prompts. The authors also annotate 518 language prompts for five existing night UAV tracking datasets to enable vision-language evaluation. They report state-of-the-art AUC and mACC on these benchmarks, with 2.8x higher speed and 50.2% lower GPU memory than CiteTracker, and they release code. The internal ablations show gains from both MLLE and CMM, and the efficiency numbers are concrete, but the headline SOTA comparison is complicated by the fact that MambaTrack is the only method evaluated with per-video language prompts while the compared baselines are cited from their original papers.","tokens_in":1498,"tokens_out":2390,"duration_ms":80940,"significance":"If the results hold, MambaTrack would be a valuable efficient baseline for night UAV tracking, showing that a mamba-based architecture can compete with transformer-based vision-language trackers while using far fewer learnable parameters and less GPU memory. The language-prompt annotations could also serve as a useful resource for future vision-language tracking research. The paper provides a code link, component-wise ablations in Table I, and concrete efficiency numbers in Table III, which are strengths. However, the central SOTA claim is currently not established because the evaluation protocol gives MambaTrack an extra test-time language input that the compared trackers were not given, as shown by the ablation dependence on the CMM module.","major_comments":[{"comment":"The SOTA comparison is not apples-to-apples. In §III-A the authors state that they manually annotate 518 language prompts for the night UAV datasets, and in §III-B they state that they primarily adopt tracking results of methods from original datasets. MambaTrack is therefore the only method in Figs. 3-4 that receives per-video natural-language input on these benchmarks. Table I shows that this extra input is load-bearing: on UAVDark135, Baseline+MLLE (no language) achieves 54.5 AUC, below CiteTracker's reported 55.3, and the CMM module with language prompts contributes the +2.7 AUC that lifts MambaTrack to 57.2. Since CiteTracker and JointNLT are themselves vision-language trackers, they would plausibly benefit from the same prompts. The headline claim of best AUC and mACC is not supported until the compared VL trackers are re-run under the same prompt-conditioned protocol, or a language-free comparison is reported.","section":"§III-A, §III-B, Table I, Figs. 3-4"},{"comment":"The training protocol for the language branch is unspecified. The paper states that the tracker is trained on TrackingNet, GOT-10k, LaSOT, COCO, and WebUAV-3M, but does not state whether these datasets provide language descriptions for the visual grounding part, how the CMM network is trained with paired vision-language inputs, or whether the language encoder (Mamba-130M) is frozen. Without this information, the ablation gain of the CMM in Table I cannot be interpreted, and there is a risk that the language branch is only exposed to prompts at test time. Please specify the text data used during training and the freezing or fine-tuning of the language encoder.","section":"§III-A"},{"comment":"The efficiency comparison is under-specified. The table reports speed and GPU memory on a single RTX A6000, but the paper does not say whether the language encoder is executed once per video or per frame, which materially affects FPS and memory numbers. It also does not state whether the values for STARK50, JointNLT, VLT_TT, and CiteTracker were re-measured in the same software environment or taken from their original papers. The 2.8x faster and 50.2% memory reduction claims require this measurement consistency.","section":"Table III, §III-D"}],"minor_comments":[{"comment":"The symbols f_m and f_v in Eqs. (6) and (7) are not defined; they presumably refer to h_m and h_v from Eq. (5), or to a different feature map. Please define all variables used in these equations.","section":"Section II-C, Eqs. (6)-(7)"},{"comment":"The header P P norm is not expanded; please define P and P_norm in the table caption or in the text.","section":"Table I"},{"comment":"Please state whether the low-light enhancer (MLLE) is fine-tuned during tracker training or kept frozen after being trained on LOL.","section":"§III-A"},{"comment":"The GitHub link at the end of the abstract points to Awesome-Multimodal-Object-Tracking, which appears to be a collection or survey rather than a dedicated code repository for MambaTrack; please confirm that the link contains the code used in the paper.","section":"Abstract / Code availability"},{"comment":"The phrase the first mamba-based baseline tracker for night UAV tracking is a strong novelty claim; consider softening it to a mamba-based tracker unless a comprehensive prior-art search supports the uniqueness.","section":"Introduction, Contributions"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the evaluation asymmetry: MambaTrack receives per-video language prompts that the cited baselines were not given, and the ablation shows that this extra supervision is what lifts the method above CiteTracker on UAVDark135. If the authors cannot re-run the compared vision-language trackers with the same prompts, they should either downgrade the SOTA claim to a more limited claim or report a language-free version of MambaTrack as the primary comparison. The paper also needs to clarify the training data and protocol for the language branch, as well as the speed and memory measurement details. The code link and ablation structure are positive, but the current manuscript does not sufficiently support the headline performance claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent engineering paper with a real efficiency win, but its headline performance claim is not yet established because the comparison gives MambaTrack language prompts that no baseline receives.\n\nThe genuinely new pieces are the first mamba-based tracker for night UAV tracking, a clean combination of a Retinex-style mamba enhancer (borrowed from RetinexMamba/Retinexformer) with a cross-modal gating block (from Mamba-FETrack/Pan-Mamba), and a new manually annotated set of 518 language prompts over five existing benchmarks. The ablations are informative: adding the enhancer gives +2.7/+3.1 AUC on NAT2021/UAVDark135, and adding the language branch gives another +1.8/+2.7. The efficiency numbers are concrete and attractive: 42 FPS vs CiteTracker's 11, 15.9M learnable params, 50.2% less GPU memory. That part holds up.\n\nThe soft spot is the SOTA comparison. The paper says they 'primarily adopt tracking results of methods from original datasets,' so CiteTracker, JointNLT, and the rest were evaluated without language prompts, while MambaTrack gets a hand-written sentence per video. That is not a minor protocol detail. On UAVDark135, baseline+MLLE alone scores 54.5 AUC, below CiteTracker's reported 55.3; the language branch is exactly what pushes MambaTrack to 57.2. Same picture on NAT2021. Since the compared methods include VL trackers, they would reasonably benefit from the same prompts. Without a re-run under a matched protocol, the 'best AUC' claim is not supported. The paper does release the prompts, which is good, and a revised version could fix this by evaluating baselines with prompts too.\n\nAlso minor: the GitHub link in the abstract points to an 'Awesome-Multimodal-Object-Tracking' list, not to actual code. That should be corrected.\n\nThis is not a broken paper. The architecture is sensible, the ablations are honest, and the efficiency advantage is real. The performance claim is conditional on a fair comparison. I'd send it to review, but with a clear request to redo the SOTA table under the same prompt-conditioned protocol. For anyone working on mamba trackers or night UAV tracking, this is a useful baseline, but I wouldn't cite the SOTA numbers as they stand.","headline":"Solid mamba tracker with real efficiency gains, but the SOTA claim depends on language prompts that baselines never saw—so take the head-to-head numbers with a grain of salt.","tokens_in":10962,"tokens_out":3088,"would_cite":false,"duration_ms":25816,"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":"MambaTrack, a mamba-based tracker with low-light and language enhancement, claims state-of-the-art accuracy on five night UAV tracking benchmarks while running 2.8x faster than the transformer-based CiteTracker and using 50.2% less GPU…","keywords":["night UAV tracking","mamba network","low-light enhancement","vision-language tracking","state space model","cross-modal fusion","Retinex theory","efficient tracking"],"falsifier":"Re-run MambaTrack on the five benchmarks with the language branch disabled (visual-only) and compare against CiteTracker and STARK50; if the AUC drops below CiteTracker's, the reported SOTA result depends on the test-time language prompts rather than the proposed architecture. Alternatively, give the same per-video language prompts to the compared trackers (e.g., by fine-tuning them on the annotated dataset) and see if the gap persists.","tokens_in":9988,"feed_emoji":"🌙","tokens_out":6446,"duration_ms":49644,"temperature":0.7,"pith_summary":"MambaTrack proposes the first mamba-based tracker for night UAV tracking, combining a mamba-based low-light enhancer (MLLE) with a cross-modal mamba (CMM) network that fuses visual features with language prompts. The paper claims this dual enhancement achieves the highest AUC and mACC scores among all compared trackers on five night UAV tracking benchmarks, while running 2.8× faster than the transformer-based CiteTracker and using 50.2% less GPU memory. The authors further introduce a new vision-language night UAV tracking task by manually annotating 518 language prompts across the existing datasets. If correct, the work shows that state-space-model architectures can match or beat transformer trackers in low-light aerial tracking at a fraction of the computational cost.","feed_headline":"Mamba tracker tops night-UAV benchmarks, 2.8x faster","feed_subtitle":"Dual enhancement—low light and language—raises AUC to 57.2 on UAVDark135 while halving GPU memory.","key_machinery":"The load-bearing objects are the mamba-based low-light enhancer (MLLE) and the cross-modal mamba (CMM) network. MLLE operationalizes Retinex decomposition with an illumination estimator ($IE$) that outputs a light-up map $\\bar{L}$ and feature map $F_{lu}$, plus a damage restorer ($DR$) built from illumination fusion state space models (IFSSM) that removes the corruption term $C = R\\otimes(\\hat{L}\\otimes\\bar{L}) + \\hat{R}\\otimes(L+\\hat{L})\\otimes\\bar{L}$, yielding $I_{en} = I\\otimes\\bar{L} + DR(I\\otimes\\bar{L}, F_{lu})$. CMM maps visual embeddings $H_x$, $H_z$ and language embedding $H_t$ into a shared space, computes $y_m = SSM(SiLU(Conv(f_m)))$ and gates with $z_m = y_m \\otimes SiLU(f_v)$, then adds the gated branches to produce language-enhanced search and template embeddings. Together they give the tracker global low-light enhancement with local detail preservation and semantic fusion, all under mamba's linear computational complexity.","core_discovery":"On the paper's own terms, the central discovery is that a tracker built on mamba's linear-time state space model, with a Retinex-based low-light enhancer and a language-guided cross-modal gating network, outperforms all compared visual-only and vision-language trackers on DarkTrack2021, NAT2021, NAT2021L, UAVDark70, and UAVDark135. The low-light enhancer decomposes the image as $I = (R+\\hat{R})\\otimes(L+\\hat{L})$ per Retinex theory, estimates a light-up map $\\bar{L}$ satisfying $\\bar{L}\\otimes L=1$, and produces $I_{en} = I\\otimes\\bar{L} + DR(I\\otimes\\bar{L}, F_{lu})$, where the damage restorer $DR$ removes noise and artifacts. The cross-modal mamba network injects language embeddings into search embeddings via gating and state space models, yielding language-enhanced features that sharpen the response map. The paper reports the best AUC of 57.2% on UAVDark135 with only 15.9M learnable parameters and 42 FPS inference, framing mamba as a more efficient backbone than transformers for night UAV tracking.","pith_inferences":["In our view, a key fairness question the paper leaves open is whether the language prompts, written per test video, constitute extra test-time supervision; the SOTA comparison may be less clean if other trackers are evaluated without equivalent semantic hints.","Our reading of the ablation is that the low-light enhancer alone accounts for most of the improvement, so the architecture's core advantage might be in image enhancement rather than multimodal fusion; the language branch's marginal gain is smaller.","Because each benchmark has only 23–180 videos, the reported AUC gaps of 1–2 points could be within statistical noise; a permutation test across frames or repeated evaluation with different prompt phrasings would clarify whether the gains are robust.","We speculate that the same dual-enhancement recipe could transfer to other low-light vision tasks, but that extension is not explored in the paper."],"forward_implications":["MambaTrack's efficiency numbers (42 FPS, 1142 MB GPU memory, 15.9M learnable parameters) suggest mamba-based trackers are viable for real-time UAV deployment, where transformer trackers may be too heavy.","The 518 manually annotated language prompts create a new vision-language night tracking benchmark task, available to the community once released.","The ablation study attributes 2.7–3.1% AUC gains to MLLE and 1.8–2.7% to CMM, indicating that low-light enhancement is the larger contributor to night tracking performance.","The attribute-based evaluation shows MambaTrack leads on fast motion, illumination variation, low resolution, occlusion, and viewpoint change, covering the main failure modes of night UAV tracking."],"supporting_citations":[{"why":"Provides the linear-time state space model (Mamba) backbone that gives MambaTrack its efficiency advantage.","marker":"[16]"},{"why":"Supplies the Retinex decomposition and perturbation modeling (Eq. 1) that MLLE adapts for low-light enhancement.","marker":"[19]"},{"why":"CiteTracker is the main vision-language SOTA baseline on UAVDark135 that MambaTrack claims to beat in accuracy (57.2 vs 55.3 AUC) and efficiency (2.8x faster, 50.2% less GPU memory).","marker":"[15]"},{"why":"JointNLT is the vision-language tracking baseline whose language-prompt integration approach MambaTrack follows for adding the [CLS] token and language embeddings.","marker":"[14]"},{"why":"Provides the NAT2021 benchmark and its unsupervised domain adaptation baseline, two of the five night tracking test sets and the night-domain challenges.","marker":"[4]"},{"why":"The UAVDark135 all-day tracking benchmark is the primary dataset for the headline AUC comparison and attribute-based evaluation.","marker":"[2]"},{"why":"The DarkTrack2021 dataset and its transformer enhancer represent both a test set and a prior low-light enhancement approach.","marker":"[10]"},{"why":"ADTrack is the pioneering anti-darkness UAV tracker whose tone-mapping approach MLLE contrasts with.","marker":"[11]"}],"fun_headline_variants":["Night UAV tracker: mamba + dual enhancement, 2.8x faster","2.8x faster: mamba night UAV tracker with dual boost","Dual-enhancement mamba tracker: night UAV, 2.8x faster","Mamba-based night UAV tracker: 2.8x faster, halves GPU","Night UAV tracking: mamba dual-enhancement beats all, 2.8x speedup"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison assumes that giving MambaTrack a manually written language prompt for every test video is not extra test-time supervision that the compared trackers lack, so the SOTA accuracy claim may be unfair.","fun_headline_variants_meta":{"raw":{"variants":["Night UAV tracker: mamba + dual enhancement, 2.8x faster","2.8x faster: mamba night UAV tracker with dual boost","Dual-enhancement mamba tracker: night UAV, 2.8x faster","Mamba-based night UAV tracker: 2.8x faster, halves GPU","Night UAV tracking: mamba dual-enhancement beats all, 2.8x speedup"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001373,"raw_usage":{"total_tokens":5574,"prompt_tokens":968,"completion_tokens":4606,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":584,"completion_tokens_details":{"reasoning_tokens":4494}},"tokens_in":584,"tokens_out":4606,"duration_ms":30537,"temperature":1.0,"reasoning_tokens":4494,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T13:55:24.905016+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run MambaTrack on the five benchmarks with the language branch disabled (visual-only) and compare against CiteTracker and STARK50; if the AUC drops below CiteTracker's, the reported SOTA result depends on the test-time language prompts rather than the proposed architecture. Alternatively, give the same per-video language prompts to the compared trackers (e.g., by fine-tuning them on the annotated dataset) and see if the gap persists.","supporting_citations":[{"cited_title":"Citetracker: Correlating image and text for visual tracking,","cited_arxiv_id":null,"evidence_quote":"CiteTracker is the main vision-language SOTA baseline on UAVDark135 that MambaTrack claims to beat in accuracy (57.2 vs 55.3 AUC) and efficiency (2.8x faster, 50.2% less GPU memory)."},{"cited_title":"Unsupervised domain adaptation for nighttime aerial tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the NAT2021 benchmark and its unsupervised domain adaptation baseline, two of the five night tracking test sets and the night-domain challenges."},{"cited_title":"Adtrack: Target-aware dual filter learning for real-time anti-dark uav tracking,","cited_arxiv_id":null,"evidence_quote":"ADTrack is the pioneering anti-darkness UAV tracker whose tone-mapping approach MLLE contrasts with."}],"review_version":1}