{"id":"3ebb8c32-dce1-4682-8046-7ee1cfa7cd07","arxiv_id":"2504.13604","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"FocusTrack adaptively enlarges a local tracker's search region when its confidence drops, and uses attention-generated masks to refine features, reaching 67.7% AUC on AntiUAV and 62.8% on AntiUAV410 at real-time speeds.","lead":"FocusTrack is a drone-tracking algorithm that automatically widens its search window when the drone disappears from view, and it runs fast enough for real-time use. It improves tracking accuracy on two infrared drone benchmarks while using far less computation than global-search trackers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"SRA's presence classifier is trained on cross-video negative pairs, so it may learn video identity rather than target absence; the adaptive-expansion mechanism—the core of the claimed contribution—may not trigger on same-video out-of-view events.","rationale":"The most load-bearing step in the paper is the transition from training SRA to its inference role. The abstract's accuracy claim depends on the 2–9% AUC gains over the static-search baseline, and Table IV attributes those gains to SRA. But SRA's only sensor is the CLS token, and the contrastive sampling in Section III-B2 trains that token to distinguish same-video from different-video pairs. At inference the hard cases are same-video pairs where the target has left the crop. This is a train/test task mismatch, not a disagreement with community consensus. It is likewise not resolved by the attribute-based success plots: those measure end-to-end AUC, which could improve for reasons other than true absence detection (e.g., occasional low-confidence frames, or the ATM module). The concrete logit-separation test would settle it. If the logits separate well, the concern is answered and the CONDITIONAL verdict can stand. If they do not, the central mechanism of the paper is not doing the work attributed to it, and the paper would need to be revised to either change the negative sampling or temper the claim. I agree with the reader's weakest_assumption; my proposed test is a sharper operationalization of it. The paper does make independent contributions (released code, consistent retraining protocol, ablations), but none of these directly validates the SRA sensing mechanism, which is why the concern remains load-bearing.","tokens_in":20770,"tokens_out":5707,"duration_ms":51902,"concrete_test":"Compute SRA logits on held-out AntiUAV410 test frames where the ground-truth target center lies outside the current search region (target-absent crops) and on frames where it lies inside (target-present crops), using the released code and the paper's inference settings. If the AUC of these logits as a binary classifier is below roughly 0.8, or if the target-absent logits rarely fall below Tlogits=0.8, then the SRA module is not detecting target absence as claimed and the adaptive-expansion mechanism is not responsible for the reported gains.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that SRA makes a local tracker competitive with global trackers by expanding the search region when the target leaves the field of view. For this to hold, the CLS-token logits must indicate 'target absent from the current search crop.' Section III-B2 defines negative pairs as frames from different sequences (presence=0) and positive pairs as frames from the same sequence (presence=1). At inference the template and search are always from the same video; an abrupt camera motion or out-of-view event leaves the crop target-free but the two inputs are still same-video. The training signal therefore supervises sequence identity or global appearance agreement, not the presence of the target inside the crop. A model can satisfy the contrastive objective while being blind to the actual failure mode it must detect, and the qualitative SRA visualization in Fig. 10 cannot distinguish the two cues. If the logits do not drop when the target exits, Algorithm 1 never expands (or expands only on incidental appearance changes), and the tracker has no recovery mechanism beyond fmax=8. A second, related concern is that the operating thresholds Tlogits=0.8, Tscore=0.5, fmax=8, fstep=1 are all selected on the test set via Table VI, so the reported 62.8% AUC is in part a selected result. Both issues bear on the same central mechanism and are not addressed by the paper.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FocusTrack, a local anti-UAV tracker built on OSTrack, adding two modules: Search Region Adjustment (SRA), which uses a CLS token to estimate a target-presence probability and adaptively expands the search crop during inference when confidence is low; and Attention-to-Mask (ATM), which aggregates search features from multiple backbone layers via cross-attention to produce a mask that refines the classification score map. The method is trained on the AntiUAV410 training set and evaluated on both AntiUAV and AntiUAV410, reporting 67.7% and 62.8% AUC, respectively, and outperforming retrained local baselines by 8.5-9.1 AUC points. The paper also reports 30G MACs, 143 fps for the SRA-only variant, and 44 fps for the full version, and includes extensive ablations and attribute-based analyses.","tokens_in":21074,"tokens_out":5457,"duration_ms":47561,"significance":"If the reported numbers hold, FocusTrack offers a credible efficiency-accuracy trade-off for anti-UAV tracking: the SRA-only variant is an order of magnitude cheaper than global redetection trackers, and the full version improves over strong local baselines by a substantial margin. The retraining of all local baselines on the same training set is a fairer comparison than relying on published numbers, and the module ablations in Tables IV-VIII are internally consistent. However, the headline 'state-of-the-art' claim is not supported by Table III, where the global tracker SiamDT reaches 66.8% AUC versus FocusTrack's 62.8% AUC, and the core generalization claims rest on two methodological choices that need scrutiny: the contrastive training signal for the presence classifier and the selection of inference thresholds on the test set.","major_comments":[{"comment":"The negative pairs for the SRA presence classifier are defined as frames from different sequences, while the inference-time failure the classifier is meant to detect is a target-absent search crop from the same sequence (e.g., abrupt camera motion or out-of-view events). A cross-video negative pair can be solved by global appearance or sequence-identity cues, so the CLS logits are not necessarily evidence that the target is absent from the current search crop. The qualitative visualization in Fig. 10 cannot separate these cues. This is load-bearing because Algorithm 1 expands the search region only when both logits<T_logits and Pmax<T_score. Please add same-video negative pairs in which the target is outside the search crop, and report an ablation or per-sequence logits around out-of-view events showing that logits drop specifically for same-video target absence.","section":"III-B2"},{"comment":"The four inference hyperparameters Tlogits=0.8, Tscore=0.5, fmax=8, and fstep=1 are selected by sweeping on the AntiUAV410 test set in Table VI, and no separate validation split or repeated runs are reported. The final 62.8% AUC is therefore in part a test-set-selected number, and the 0.2-point difference between the best Table VI configuration (62.6%) and the final two-phase result (62.8%) is within the range that could be noise. Please tune the thresholds on the 90-sequence validation split or via cross-validation, report test numbers for the validation-tuned configuration, and provide error bars over multiple runs.","section":"IV-A4, Table VI"},{"comment":"The paper claims 'state-of-the-art performance,' but Table III reports SiamDT at 66.8% AUC and 90.0% precision on AntiUAV410 versus FocusTrack's 62.8% AUC and 86.2% precision. The correct claim is that FocusTrack is state-of-the-art among local trackers and offers a much better efficiency-accuracy trade-off than the global tracker. Please revise the abstract and conclusion to avoid an unsupported global-state-of-the-art claim.","section":"Abstract, Conclusion, Table III"}],"minor_comments":[{"comment":"The table header contains a typo: 'SPEES' should be 'SPEED'.","section":"Table III"},{"comment":"The sentence 'the lower section presents results from models retrained on AntiUAV410 test set' should read 'training set', since the lower part of Table II is explicitly reported as retrained on the AntiUAV410 training set.","section":"IV-B2"},{"comment":"The caption contains a typo: 'traget motion' should be 'target motion'.","section":"Fig. 1 caption"},{"comment":"The section heading 'Attention-to Mask Module' should be hyphenated as 'Attention-to-Mask Module' for consistency with the text.","section":"III-C"},{"comment":"In the Single-Phase Training description, the text says '60,000 positive samples per iteration' and then derives 85,715 total samples; please clarify whether these are per-iteration samples or per-epoch samples, and define the epoch/iteration relationship.","section":"IV-A3"}],"recommendation":"major_revision","confidential_remarks":"The two main technical risks are the SRA training signal (cross-video negatives may not teach same-video target absence) and the test-set-driven threshold selection. Both are fixable with additional experiments, and the paper otherwise contains a thorough, fair comparison with local baselines. I would encourage the editor to require same-video negative-pair experiments and validation-based threshold reporting before publication; without them, the central adaptive-expansion mechanism is not fully established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a legitimate, reasonably careful engineering paper that deserves proper peer review, but not a clean accept as-is. The new thing is the SRA scheduler—confidence-triggered search-factor expansion for a local transformer tracker—plus the ATM mask refinement ported from SegViT. The gains over a fixed search factor are real: simply widening OSTrack's search from 4 to 6 gives 6.1 AUC on AntiUAV410, and SRA adds another 2.1. That is a meaningful efficiency-accuracy trade for anti-UAV tracking.\n\nCredit where due: they retrain all competing local trackers on the same training set, report MACs and fps on the same GPU, and release code. The ablations are internally consistent, and the paper honestly shows that the largest single gain comes from a larger fixed search region, not from the adaptive mechanism.\n\nSoft spots, in order of seriousness. First, the SRA presence classifier is trained with negative pairs from different videos, so the contrastive supervision encodes sequence identity, not target absence in a same-video crop. The stress-test raised this, and I think it lands. The Pmax threshold partially papers over it because the classification peak also gates expansion, but the paper never analyzes whether the CLS logits actually drop when the target leaves a same-video crop. The qualitative Fig. 10 cannot distinguish the two cues. Second, all four inference thresholds (Tlogits, Tscore, fmax, fstep) are selected by sweeping on the test set (Table VI). That makes the reported degree of improvement partly a selection artifact; there is no validation split and no error bars. Third, the \"state-of-the-art\" claim is overbroad: SiamDT gets 66.8 AUC on AntiUAV410, 4 points higher, at 8 fps. FocusTrack's point is the efficiency-accuracy trade, not absolute SOTA.\n\nNone of this is fatal. The core claim—a local tracker with adaptive search can get within 4 AUC of a global tracker at 5x speed—holds up as an engineering result. The paper would be stronger if they retrained SRA with same-video target-absent negatives, fixed thresholds on validation, and reported standard errors. Who is it for? People working on anti-UAV or long-term tracking with abrupt motion. I would cite it. It deserves refereeing—the claims are concrete, artifacts are available, and the core idea is worth scrutiny. I would accept it conditionally on addressing the SRA training mismatch and moving threshold selection off the test set.","headline":"Solid, reproducible engineering on a hard benchmark; the adaptive search mechanism is real but its presence classifier is trained on the wrong contrast, and the headline numbers lean on test-set-tuned thresholds.","tokens_in":21625,"tokens_out":2595,"would_cite":true,"duration_ms":24076,"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":"FocusTrack claims that a local tracker with a self-adaptive search region and mask-refined features matches global trackers on anti-UAV benchmarks while costing a fraction of the compute.","keywords":["anti-UAV tracking","single object tracking","thermal infrared","vision transformer","adaptive search region","attention-to-mask","small target tracking"],"falsifier":"Run FocusTrack on a held-out set of unseen anti-UAV sequences with extreme camera shakes or adversarially selected jumps that place the target outside a factor-8 window, and record whether the presence logits stay below 0.8 while the search factor climbs to its maximum without reacquiring the target. A concrete failure would be long runs where the classifier keeps the window small while the target is outside it, or where expansion reaches $f_{max}=8$ and the target is never recovered.","tokens_in":20550,"feed_emoji":"🎯","tokens_out":7014,"duration_ms":60831,"temperature":0.7,"pith_summary":"FocusTrack tries to settle a trade-off in anti-UAV tracking: local trackers are fast but lose small drones when the camera jerks, while global trackers that scan the whole frame rarely lose the target but are too slow for real-time use. The paper argues that a local tracker can get most of the global tracker's robustness by making its search-region size responsive to an estimated probability that the target is still in view, and by using a mask-based attention module to keep the target representation sharp when the field of view widens. On the thermal-infrared AntiUAV and AntiUAV410 benchmarks, FocusTrack reports 67.7% and 62.8% AUC, 8.5 and 9.1 points above its OSTrack baseline. If the result holds, a tracker that costs only 30G MACs and runs at 44-143 fps can replace global re-detection trackers in drone-defense applications.","feed_headline":"Adaptive search window lifts local UAV tracker past global ones","feed_subtitle":"Self-adjusting search regions and attention masks deliver 67.7% AUC on AntiUAV at roughly one-eighth the compute of global trackers.","key_machinery":"The load-bearing mechanism is the pair (SRA, ATM). SRA is an adaptive search-region controller: a learnable CLS token is prepended to the concatenated template/search tokens of the ViT backbone, trained with cross-entropy on positive pairs from the same video and negative pairs from different videos, and at inference its logits, combined with the peak score of the classification map, decide whether the search factor grows by a step of 1 up to a maximum of 8 (thresholds $T_{logits}=0.8$ and $T_{score}=0.5$). ATM is a stack of three cross-attention blocks that convert a learnable query into sigmoid-activated masks from backbone layers 6, 8 and 12; the fused mask re-weights the classification map and is supervised by focal loss on rectangular pseudo-masks derived from bounding boxes. Together they keep the target inside the field of view while preventing a larger view from diluting target features.","core_discovery":"FocusTrack's central claim is that search-region size, not feature strength alone, is the main failure point for local trackers on anti-UAV videos, and that it can be made adaptive without paying a global-search cost. Its Search Region Adjustment (SRA) module prepends a CLS token to the template and search tokens of a ViT-based tracker, trains that token with contrastive positive/negative frame pairs, and then uses its presence logits plus the peak classification score to decide whether to keep the search factor at 6 or expand it stepwise up to 8. The Attention-to-Mask (ATM) module counteracts the feature dilution caused by larger fields of view by generating hierarchical segmentation-mask attention that sharpens the classification map. With both modules, the paper reports state-of-the-art success rates on AntiUAV (67.7% AUC) and AntiUAV410 (62.8% AUC), surpassing local baselines and coming within 4.0 AUC of the global tracker SiamDT while using roughly one-eighth of its compute.","pith_inferences":["My inference: the fixed thresholds $T_{logits}=0.8$ and $T_{score}=0.5$ were tuned on test-set sweeps, so a real deployment would need online calibration because a motion pattern that fools the presence classifier will exhaust the $f_{max}=8$ expansion and permanently lose the target.","My inference: because the presence classifier is trained on pairs drawn from the same benchmark, it may be sensitive to domain shift in camera motion; a testable extension is to train SRA on motion-perturbed frames with pseudo-presence labels.","My inference: ATM's mask output opens a path toward joint tracking and segmentation-based small-target detection, and replacing rectangular pseudo-masks with finer annotations could improve performance on tiny-detail cases."],"forward_implications":["If FocusTrack's reported gains hold, local trackers can close most of the accuracy gap to global re-detection trackers on anti-UAV benchmarks at a fraction of the compute (30G MACs versus 225.3G for SiamDT).","The ablation shows that raising the fixed search factor from 4 to 6 adds 6.1 AUC and the adaptive SRA adds another 2.5 AUC, meaning search-window sizing is a first-order error source for small-target tracking, not a minor tuning detail.","The full FocusTrack stays real-time at 44 fps while also producing segmentation masks; dropping the ATM module still yields 62.3% AUC at 143 fps, so the adaptive window alone carries most of the benefit."],"supporting_citations":[{"why":"Supplies the OSTrack one-stream transformer baseline, the fixed search-factor crop convention, and the 53.7% AUC starting point for all ablations.","marker":"[3]"},{"why":"DropMAE pretraining initializes the ViT backbone; the ablation credits it with a 0.4 AUC gain over random initialization.","marker":"[9]"},{"why":"Provides the AntiUAV410 benchmark and the SiamDT global tracker used as the accuracy and compute reference.","marker":"[1]"},{"why":"Provides the AntiUAV benchmark and the success-rate, precision, normalized precision, and state-accuracy metrics used throughout the evaluation.","marker":"[5]"},{"why":"SegViT supplies the attention-to-mask design that the ATM module adapts into stacked cross-attention blocks with a learnable class token.","marker":"[12]"},{"why":"ViT provides the patch embedding, CLS token, and positional embedding machinery used by both the backbone and the SRA module.","marker":"[24]"}],"fun_headline_variants":["Self-adaptive search window lifts local UAV tracker to near-global AUC","Adaptive search region closes the gap to global trackers on anti-UAV","SRA and ATM: local tracker with near-global accuracy at 8x less compute","Adaptive search region keeps UAV targets locked at low compute","FocusTrack: real-time anti-UAV tracking with self-adjusting view"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the presence classifier and its two fixed thresholds: if the CLS token fails to notice that the target has left the view, or the chosen thresholds do not transfer to new motion patterns, the search region never expands and the tracker has no way to recover the target.","fun_headline_variants_meta":{"raw":{"variants":["Self-adaptive search window lifts local UAV tracker to near-global AUC","Adaptive search region closes the gap to global trackers on anti-UAV","SRA and ATM: local tracker with near-global accuracy at 8x less compute","Adaptive search region keeps UAV targets locked at low compute","FocusTrack: real-time anti-UAV tracking with self-adjusting view"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00114,"raw_usage":{"total_tokens":4807,"prompt_tokens":1094,"completion_tokens":3713,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":710,"completion_tokens_details":{"reasoning_tokens":3624}},"tokens_in":710,"tokens_out":3713,"duration_ms":24158,"temperature":1.0,"reasoning_tokens":3624,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:03:40.508831+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run FocusTrack on a held-out set of unseen anti-UAV sequences with extreme camera shakes or adversarially selected jumps that place the target outside a factor-8 window, and record whether the presence logits stay below 0.8 while the search factor climbs to its maximum without reacquiring the target. A concrete failure would be long runs where the classifier keeps the window small while the target is outside it, or where expansion reaches $f_{max}=8$ and the target is never recovered.","supporting_citations":[{"cited_title":"Dropmae: Masked autoencoders with spatial-attention dropout for tracking tasks,","cited_arxiv_id":null,"evidence_quote":"DropMAE pretraining initializes the ViT backbone; the ablation credits it with a 0.4 AUC gain over random initialization."},{"cited_title":"Anti-uav410: A thermal infrared benchmark and customized scheme for tracking drones in the wild,","cited_arxiv_id":null,"evidence_quote":"Provides the AntiUAV410 benchmark and the SiamDT global tracker used as the accuracy and compute reference."},{"cited_title":"Anti-uav: a large-scale benchmark for vision-based uav tracking,","cited_arxiv_id":null,"evidence_quote":"Provides the AntiUAV benchmark and the success-rate, precision, normalized precision, and state-accuracy metrics used throughout the evaluation."},{"cited_title":"Segvit: Semantic segmentation with plain vision transformers,","cited_arxiv_id":null,"evidence_quote":"SegViT supplies the attention-to-mask design that the ATM module adapts into stacked cross-attention blocks with a learnable class token."}],"review_version":1}