{"id":"c9bac9f1-8851-44ad-a7ec-ebb62f77bbd1","arxiv_id":"2505.00752","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"DARTer combines a Dynamic Feature Blender and a Dynamic Feature Activator to improve nighttime UAV tracking accuracy and efficiency, outperforming prior trackers on five benchmarks.","lead":"A new nighttime UAV tracker, DARTer, fuses static and dynamic template features with cross-attention and dynamically skips ViT layers to save computation, reporting state-of-the-art results on five benchmarks. The method is efficient at 74 frames per second, but it lacks released code and a clear training procedure for its discrete gating mechanism.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"DFA hard-threshold gate has no training path; without an undisclosed relaxation, the gate weights cannot be learned and the claimed DFA gains are unexplained.","rationale":"The reader's weakest assumption is exactly the DFA training gap, and I find it decisive for the central claim as stated. Section 2.3 explicitly says no complex multi-task loss is needed, which rules out an unmentioned auxiliary gating loss unless the paper is incomplete. The hard threshold in Section 2.2 has zero gradient almost everywhere, so the DFA parameters cannot be learned through the published loss. The Table 4 ablation makes this load-bearing: DFA alone contributes 1.8 precision points, so the headline improvement depends on the gate being learned. Secondary issues include consistency slips (Section 3.2 refers to 'DFTrack'; Table 3's header says 'NAT2021-L'), the absence of code, and no variance estimates, but these do not change the main verdict. The proposed gradient-norm check would settle the concern cleanly: if gradients are nonzero, the authors should disclose the relaxation; if zero, the reported DFA gains cannot be reproduced as described. The empirical comparison may still be valuable, so the conditional verdict remains appropriate.","tokens_in":9625,"tokens_out":4089,"duration_ms":45167,"concrete_test":"Instrument the training loop and log the L2 norm of gradients for the DFA linear and Conv parameters (Eq. 4) over the first several batches. If the gradient norm is identically zero, the described end-to-end training cannot learn the gate, confirming the concern. If the norm is nonzero, the authors should identify the relaxation or estimator used, since none is stated; the test then settles whether a missing mechanism exists.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2.2 defines the activation probability p_i = sigma(L(r_i)+Conv(r_i)) and then applies a hard threshold: the i-th ViT block is executed only if p_i > beta, otherwise it is skipped. Section 2.3 gives the total loss as lambda1*L_ce + lambda2*L_SIoU and explicitly states that no complex hand-designed losses are needed. There is no straight-through estimator, Gumbel-softmax relaxation, REINFORCE estimator, or auxiliary gating loss described. Because the forward path applies a step function of p_i, the derivative of the skip decision with respect to the DFA linear/conv parameters is zero almost everywhere, so end-to-end gradient training cannot update those parameters. Table 4 attributes a precision gain of 1.8 points on NAT2024-1 to the DFA; if the gate parameters receive no gradient, that gain must come from random gating or from an undocumented training procedure. This is the load-bearing weak point: the paper's central accuracy/efficiency claim depends on DFA, yet the mechanism by which DFA is optimized is absent from the manuscript.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents DARTer, a single-stream transformer-based tracker for nighttime UAV tracking. The method introduces a Dynamic Feature Blender (DFB) that fuses static and dynamic template features via cross-attention, and a Dynamic Feature Activator (DFA) that skips Vision Transformer blocks based on an activation probability derived from the current tokens. Training uses a weighted combination of softmax cross-entropy and SIoU losses. DARTer is evaluated on five nighttime benchmarks and reported to outperform previous state-of-the-art methods, with a 6.3% precision improvement on NAT2021-L and a speed of over 74 FPS with 80.9M parameters.","tokens_in":9826,"tokens_out":6462,"duration_ms":57101,"significance":"If the reported results are reproducible, DARTer would provide a useful accuracy-efficiency trade-off for nighttime UAV tracking, with consistent improvements over a wide range of baselines on five benchmarks. The paper also provides an ablation of both proposed components and demonstrates real-time inference. The main significance is conditional on resolving the ambiguous training procedure for the DFA; until then, the attribution of the improvement to the adaptive gating mechanism is not supported by the manuscript as written.","major_comments":[{"comment":"The DFA's hard decision (execute the i-th ViT block only if p_i > beta, otherwise skip) is non-differentiable. The total loss in Section 2.3 is a function only of the final prediction head, and no straight-through estimator, Gumbel-softmax relaxation, REINFORCE baseline, or auxiliary gating loss is described. Consequently, the parameters of the linear layer L and convolution Conv in Eq. (4) receive zero gradient almost everywhere during end-to-end training, and the 1.8-point precision gain attributed to DFA in Table 4 cannot be explained by the stated training procedure. Please specify the exact optimization scheme for the DFA, including any relaxation or estimator, and provide an experiment showing that the gate parameters actually change during training (e.g., statistics of p_i before and after training).","section":"Section 2.2, Eq. (4) and Section 2.3"},{"comment":"The definition of the feature extraction vector v is underspecified. The text says v belongs to the standard normal distribution N(0,1), but it is not stated whether v is a learned parameter, a fixed random projection, or re-sampled at each forward pass. This matters because r_i = v * t^{i-1}(f_D) determines the activation probability, and if v is randomly sampled, the DFA input is stochastic even at test time. Please clarify the role and training status of v.","section":"Section 2.2, Eq. (4)"}],"minor_comments":[{"comment":"The header of the 'Source' column incorrectly reads 'NAT2021-L'; it should read 'DarkTrack2021'.","section":"Table 3"},{"comment":"The image sizes for search and template are stated as 128×128 and 256×256, respectively, which is the opposite of the usual convention (search region larger than template). Please confirm the correct values.","section":"Section 3.1"},{"comment":"The proposed tracker is referred to as 'DFTrack' instead of 'DARTer'.","section":"Section 3.2, first paragraph"},{"comment":"The sentence 'To avoid the situation where all blocks are not activated, we perform feature activation calculations on all blocks except the first ViT block' is ambiguous; it should simply state that the first ViT block always executes and the activation probability is computed for the remaining blocks.","section":"Section 2.2"},{"comment":"The learning rate schedule ('after 120 epochs, the learning rate decays at a rate of 10%') is ambiguous; please specify whether this is a per-epoch decay factor, a one-time decay, or a cosine schedule.","section":"Section 3.1"},{"comment":"The dynamic template update interval is said to be fixed but its value is never given; please provide it.","section":"Sections 2.2 and 3.1"},{"comment":"The text reports that the DFB increases the success score by 1.95%, but the table shows a change of 1.9 percentage points (or a relative increase of about 3.05%); please align the reported percentages.","section":"Table 4 and Section 3.3"},{"comment":"The paper uses ACM template placeholders ('Conference acronym XX', 'Woodstock NY', 2018) in the header; these should be replaced with the actual venue information.","section":"General formatting"}],"recommendation":"major_revision","confidential_remarks":"The central issue is the missing training path for the DFA. I would recommend asking the authors to clarify or provide code for the gating mechanism; without this, the paper's central contribution is not verifiable. The manuscript is formatted as an ACM conference submission with placeholders, which may indicate a venue mismatch; this is not a reason to reject but should be checked against the journal's submission guidelines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: DARTer is a plausible, incremental step for nighttime UAV tracking. The DFB (cross-attention between static and dynamic template features) is a simple and reasonable idea; the DFA (adaptive ViT layer skipping) is a known conditioning trick. The empirical evidence is the strongest part: consistent gains across five benchmarks, an ablation that shows each module contributes, and a reasonable FPS/params profile. The hand-set beta and loss weights are not fitted to the benchmarks, so there is no circularity.\n\nThe load-bearing weakness is the DFA training procedure. The paper defines an activation probability p_i, then applies a hard threshold to decide whether to execute a ViT layer. No straight-through estimator, Gumbel-softmax relaxation, or auxiliary gating loss is described. As written, the gate parameters get zero gradient from the classification and regression losses, so the 1.8-point precision gain attributed to DFA in Table 4 is unexplained. The authors cite SGDViT and AVTrack, both of which use gradient-friendly gating, so they very likely did something similar and forgot to write it down. This is fixable, but it must be disclosed.\n\nThe minor issues are copy-paste artifacts: the NAT2024-1 paragraph calls the tracker 'DFTrack', and Table 3 has a wrong header label. Easy cleanup. More substantively, there is no code and no error bars. Error bars are not standard in tracking, but the margins over DCPT and MambaNUT on some benchmarks are small, and MambaNUT gets 72 FPS with 4.1M parameters versus 80.9M for DARTer. The efficiency claim is reasonable relative to DCPT, but the framing should be more careful.\n\nWhat the paper does well: the evaluation is broader than in many tracking papers, the qualitative examples look aligned with the numbers, and the method is simple enough to be reproducible if the training details are filled in. This is a paper for the UAV tracking community, and it deserves a serious referee. I would accept it with the expectation of a major revision that adds the DFA training recipe and possibly code. Without that, the central efficiency claim is not verifiable.","headline":"Solid incremental results for nighttime UAV tracking, but the DFA gating mechanism's training recipe is missing and that gap is load-bearing.","tokens_in":10338,"tokens_out":3925,"would_cite":false,"duration_ms":38726,"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":"DARTer claims that fusing static and dynamic template features and adaptively skipping ViT layers yields state-of-the-art nighttime UAV tracking on five benchmarks, with a 6.3% precision gain on NAT2021-L at over 74 FPS.","keywords":["nighttime UAV tracking","dynamic feature fusion","adaptive computation","Vision Transformer","template matching","low-light tracking","real-time tracking"],"falsifier":"Run the exact training recipe with the DFA gating branch's gradients removed, or with the gate forced to always active, and compare NAT2021-L precision; if performance does not drop, the reported gains are not caused by adaptive layer skipping.","tokens_in":9427,"feed_emoji":"🌙","tokens_out":5761,"duration_ms":53313,"temperature":0.7,"pith_summary":"This paper is trying to establish that nighttime UAV tracking does not require a separate light enhancer or complex domain adaptation. It proposes DARTer, a single-stream tracker whose two modules, the Dynamic Feature Blender and the Dynamic Feature Activator, handle dark multi-perspective templates and skip unneeded transformer layers. The authors report state-of-the-art results across five nighttime benchmarks, including a 6.3% precision improvement on NAT2021-L, while running at over 74 frames per second. If true, this matters because it would give real-time UAV systems a way to track at night without paying the computational cost of enhancement networks.","feed_headline":"Nighttime UAV tracker tops five benchmarks at 74 FPS","feed_subtitle":"Adaptive layer skipping plus template blending beats prior state of the art on NAT2021-L by 6.3%.","key_machinery":"The load-bearing machinery is the pair of modules: DFB, a cross-attention fusion of static and dynamic template features also applied to overlapping patches, and DFA, a linear-plus-convolution gating head that maps the tokens of the previous ViT layer to a probability and skips the next layer when the probability is below the threshold 0.3. DFB makes the representation robust to viewpoint change; DFA converts fixed-depth feature extraction into adaptive computation. The single loss combining softmax cross-entropy and SIoU is deliberately simple, so the claimed gains rest on these two modules rather than on auxiliary losses.","core_discovery":"The central discovery claimed is that fusing static and dynamic template features under dark conditions, then adaptively activating only the needed ViT layers, produces both higher tracking accuracy and lower computation than prior nighttime trackers. The Dynamic Feature Blender (DFB) performs cross-attention between static and dynamic template tokens, concatenates the results, and repeats the operation on overlapped patches, producing a night-fused representation. The Dynamic Feature Activator (DFA) computes an activation probability from the previous layer's tokens and skips the next ViT block when the probability falls below 0.3, so the network spends computation only where features warrant it. With this design, DARTer reports top precision, normalized precision, and AUC on NAT2024-1, NAT2021, UAVDark135, NAT2021-L, and DarkTrack2021, reaching 85.2% precision on NAT2024-1, 64.9% precision on NAT2021-L, and over 74 FPS with 80.9M parameters.","pith_inferences":["The same static-plus-dynamic template fusion mechanism could plausibly improve daytime multi-template trackers, since nothing about DFB is inherently night-specific; the paper does not test that setting.","A natural next test is to apply DARTer's DFA to a larger ViT backbone; if the skip policy transfers, the efficiency gain should grow with depth, which the paper does not explore.","The fixed-interval dynamic template update suggests an online extension where update frequency is itself learned or scheduled, a direction the paper leaves implicit.","Because the paper does not describe how the discrete thresholding in DFA is trained, a reader should treat the speed-accuracy balance as contingent on the gating being actually learnable, not just on the reported numbers."],"forward_implications":["If the results hold, a nighttime UAV tracker can be accurate without a separate low-light enhancement network, avoiding that extra compute.","The 6.3% precision gain on NAT2021-L over the prior best would make DARTer the new reference for that benchmark's hard nighttime cases.","A 74+ FPS runtime with 80.9M parameters means the approach fits real-time use on a single consumer GPU, unlike slower SOTA trackers such as the 35 FPS darkness-prompted model it outperforms.","The ablations attribute 1.9 to 3.7 point gains on NAT2024-1 to the DFB and DFA individually, so each module carries its own contribution to the reported improvement."],"supporting_citations":[{"why":"Supplies the Overlapped ViT backbone whose patch embeddings and layers the DFB and DFA operate on.","marker":"[33]"},{"why":"Provides the NAT2021 and NAT2021-L benchmarks and the UDAT domain-adaptation baseline that DARTer surpasses.","marker":"[42]"},{"why":"Provides the UAVDark135 benchmark used to evaluate all-day and nighttime tracking.","marker":"[21]"},{"why":"Provides the DarkTrack2021 benchmark with interference-heavy nighttime sequences.","marker":"[41]"},{"why":"Is the prior darkness-prompted tracking model that DARTer reports beating on NAT2021-L and UAVDark135.","marker":"[46]"},{"why":"Is the prior Mamba-based nighttime tracker that DARTer surpasses on NAT2024-1.","marker":"[38]"},{"why":"Supplies the corner-detection-style prediction head used for classification and box regression.","marker":"[8]"},{"why":"Supplies the SIoU regression loss used in the two-term training objective.","marker":"[15]"},{"why":"Provides the cross-attention operation that the Dynamic Feature Blender uses for template fusion.","marker":"[36]"}],"fun_headline_variants":["DARTer fuses templates and skips layers to top five night tracking benchmarks","Adaptive layer skipping plus template blending: DARTer hits 74 FPS at night","Night UAV tracker blends static and dynamic templates, skips idle layers, tops five","DARTer: dynamic feature blending and activation for fast, accurate night tracking","Nighttime UAV tracker uses adaptive layer skipping to beat SOTA on five benchmarks"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole accuracy-and-efficiency story assumes the network can actually learn which transformer layers to skip, but the paper never explains how the skip decision is trained given that the threshold comparison is a hard, non-differentiable choice.","fun_headline_variants_meta":{"raw":{"variants":["DARTer fuses templates and skips layers to top five night tracking benchmarks","Adaptive layer skipping plus template blending: DARTer hits 74 FPS at night","Night UAV tracker blends static and dynamic templates, skips idle layers, tops five","DARTer: dynamic feature blending and activation for fast, accurate night tracking","Nighttime UAV tracker uses adaptive layer skipping to beat SOTA on five benchmarks"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000959,"raw_usage":{"total_tokens":4092,"prompt_tokens":957,"completion_tokens":3135,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":573,"completion_tokens_details":{"reasoning_tokens":3043}},"tokens_in":573,"tokens_out":3135,"duration_ms":22033,"temperature":1.0,"reasoning_tokens":3043,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T04:44:51.047808+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the exact training recipe with the DFA gating branch's gradients removed, or with the gate forced to always active, and compare NAT2021-L precision; if performance does not drop, the reported gains are not caused by adaptive layer skipping.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Overlapped ViT backbone whose patch embeddings and layers the DFB and DFA operate on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the NAT2021 and NAT2021-L benchmarks and the UDAT domain-adaptation baseline that DARTer surpasses."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the UAVDark135 benchmark used to evaluate all-day and nighttime tracking."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Is the prior darkness-prompted tracking model that DARTer reports beating on NAT2021-L and UAVDark135."},{"cited_title":"MambaNUT: Nighttime UAV Tracking via Mamba-based Adaptive Curriculum Learning","cited_arxiv_id":"2412.00626","evidence_quote":"Is the prior Mamba-based nighttime tracker that DARTer surpasses on NAT2024-1."}],"review_version":1}