{"id":"5e44797b-349d-4fb2-bdbf-d047e54839d0","arxiv_id":"2411.18855","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"SiamABC, a Siamese tracker with a dual-search-region, a fast filtration layer, and backward-free test-time adaptation, improves out-of-distribution tracking while running at 100 FPS on a CPU.","lead":"SiamABC is a new efficient object tracker that follows targets reliably even in fog, rain, and snow. It runs at 100 frames per second on a CPU and beats a slower, well-known tracker on a harsh out-of-distribution benchmark.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 2's dynamic-update condition is a self-comparison (`ρt > ρt`), so the printed inference procedure cannot perform the updates credited with 1.0 AUC in Table 6; the exact configuration behind the 47.2% headline is under-specified.","rationale":"The reader's weakest assumption was that the BN-statistics premise underlying DTTA transfers to tracking, with only small measured gains. I agree that the test-time components are the least secure part of the paper, but I would sharpen the concern further: the published Algorithm 2 contains an internal inconsistency that prevents the dynamic update from ever executing as written. This is a concrete correctness/reproducibility issue rather than a speculation about the BN premise. The central speed-accuracy claim is likely still supportable because the available code may implement the intended rule, and because even without the two test-time components S-Tiny's AVisT AUC would be about 0.448, still above MixFormerV2-S's 0.396. For that reason I do not recommend rejecting or moving the verdict; the appropriate outcome remains conditional, pending a code-level check of the inference loop and clarification of Eq. (7). The paper should also add the intended running-average variable to Algorithm 2 and report either repeated-run statistics or the seed sensitivity of the 1.0-1.4 AUC gains before the specific margins are taken at face value.","tokens_in":19620,"tokens_out":12481,"duration_ms":117197,"concrete_test":"Download the released S-Tiny code and compare its inference loop to Algorithm 2; then run the AVisT evaluation under three configurations: (1) exact released code, (2) released code with the dynamic-update condition disabled, and (3) a faithful implementation of the printed Algorithm 2 (which never updates because the condition is false). If configurations (1) and (3) differ by less than 0.5 AUC, the typo is benign; if they differ by roughly 2 AUC, the headline result is not reproducible from the paper text alone and the paper should be revised or conditionally accepted with code verification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline result depends on both the dynamic sample update and DTTA, but the paper's reproducible specification breaks at Algorithm 2: the update condition reads `if C ≥ N & ρt > ρt`, comparing the current classification score with itself, and no separate running-average variable is maintained even though Eq. (4) defines one. A reader implementing the printed algorithm would never update IS/ID, reverting to the 'No Updates' row of Table 6 (0.448 AUC, 2.4 points below the headline 0.472). Table 5's DTTA formula is also imprecise: Eq. (7) reuses the same symbol for the updated statistic and the current instance statistic, so the intended per-frame recurrence is not fully specified. These are not cosmetic: the two test-time components together contribute roughly 2.4 AUC (0.448 to 0.472), a substantial fraction of the 7.6-point margin over MixFormerV2-S, and the printed text does not determine how they are implemented. The released code may be correct, but the claim as written is not independently reproducible from the paper.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces SiamABC, a Siamese visual tracker with four claimed contributions: a dual-search-region representation, a lightweight Fast Mixed Filtration (FMF) layer, a transitive relation loss (TRL), and a backward-free dynamic test-time adaptation (DTTA) method. The central claim is that the smallest model, S-Tiny, achieves a 47.2% AUC on the OOD AVisT benchmark at 100 FPS on a CPU, outperforming MixFormerV2-S by 7.6% while being about 3x faster. The manuscript reports extensive experiments on 11 benchmarks, TTA baseline comparisons, and component ablations, with code and models made publicly available.","tokens_in":19896,"tokens_out":4774,"duration_ms":40505,"significance":"If the reported results hold, the paper addresses a real gap: efficient trackers that are deployable on CPU while generalizing to out-of-distribution sequences. The architectural components are coherent, and the empirical scope is broad, including 11 benchmarks and latency measurements on CPU, GPU, and an edge device. The speed-accuracy trade-off claimed for S-Tiny is attractive and largely consistent with the presented tables. The ablation study is structured and the authors provide code and models, which are strengths. The main reservations are that the printed inference-time update algorithm is not implementable as written, the 'parameter-free' characterization is contradicted by the required hyperparameters, and the OOD gains attributed to DTTA and the dynamic update are small and reported without error bars.","major_comments":[{"comment":"The printed dynamic-update procedure is not executable. In Algorithm 2, the condition 'if C >= N & rho_t > rho_t' compares the current classification score with itself, so a reader implementing the algorithm would never update I_S and I_D. Eq. (4) defines a running average but Algorithm 2 neither maintains nor compares against it; the line 'rho_t = (1 - lambda_D) rho_t + lambda_D rho_t' is also self-referential. A faithful implementation would revert to the 'No Updates' row of Table 6 (0.448 AUC), losing the +1.0 AUC attributed to this component and, together with DTTA, 2.4 AUC of the headline 0.472. Please rewrite Eq. (4) and Algorithm 2 with distinct symbols for the current score, the previous running average, and the updated running average, and state the exact configuration used for Tables 3 and 5.","section":"Section 3.2, Eq. (4), Appendix A.3 Algorithm 2"},{"comment":"The DTTA recurrence in Eq. (7) is ambiguous because the same symbols 'mu_{I,t}' and 'sigma^2_{I,t}' appear on both sides, while the text defines the right-hand side as the current instance statistics. The updated running statistics should be denoted with new symbols, and the paper should state the initialization, whether the update is applied per frame or only at update intervals, and how DTTA interacts with Algorithm 2. This matters because Table 5 credits DTTA with +1.4 AUC on AVisT (0.458 to 0.472), and the current text does not fully determine the implemented rule.","section":"Section 3.3, Eq. (7)"},{"comment":"The update strategy is repeatedly described as 'parameter-free' and 'parameterless,' but it depends on N=60 and lambda_D=0.25, and DTTA depends on lambda_BN=0.1; the training sampler also depends on Delta=150. No sensitivity analysis or validation-based selection procedure is reported for any of these values. Since these hyperparameters directly determine the OOD improvements shown in Tables 5 and 6, the 'parameter-free' claim is not supported, and the headline numbers are not tied to a reproducible selection rule. Please add sensitivity sweeps for N, lambda_D, and lambda_BN, or report a validation protocol.","section":"Section 3.2, Table 6, Section 3.3"},{"comment":"All results are single point estimates without error bars or significance testing. The main OOD improvements are small in absolute terms (DTTA +1.4 AUC on AVisT, dynamic update +1.0 AUC, S-Tiny vs. SMAT +2.5 AUC), and tracking benchmarks have substantial per-sequence variance. Please report variance over sequences or bootstrap confidence intervals for the central speed-accuracy and OOD-generalization claims so that the reader can assess whether the differences are statistically meaningful.","section":"Tables 3-6, Figure 4"}],"minor_comments":[{"comment":"The method is called 'Fast Mixed Filtration (FMT)' in the overview but 'FMF' elsewhere; please unify the abbreviation consistently.","section":"Section 3, Figure 2"},{"comment":"The benchmark name 'UA V123' should be 'UAV123', and 'A VisT' should be typeset consistently (e.g., 'AVisT') throughout the tables and text.","section":"Table 3"},{"comment":"The symbol 'P' is used for element-wise summation in the equation but resembles a parameter symbol; please use a distinct notation, such as a circled plus or an explicit 'sum', to avoid ambiguity.","section":"Eq. (1)"},{"comment":"There is a typo in 'averging in about 441 frames per sequence'; it should read 'averaging about 441 frames per sequence'.","section":"Section 4.1"},{"comment":"The comments in Algorithm 2 call N and lambda_D hyperparameters, while the surrounding text calls the strategy parameterless; please reconcile the terminology.","section":"Algorithm 2"}],"recommendation":"major_revision","confidential_remarks":"The paper claims to be the first to use test-time adaptation for single-object visual tracking. This is a strong priority claim, and I did not verify it against the broader online-adaptation-in-tracking literature; the editor may wish to ask the authors to support the claim with a more targeted citation search. I also note that the provided code was not inspected during this review; given the Algorithm 2 issue, the authors should confirm that the released code implements the intended running-average comparison and DTTA recurrence. The overall direction is promising and the empirical study is broad, so I see the issues as fixable within the manuscript's scope rather than as grounds for rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nRead SiamABC. The empirical core is real: a 2M-parameter Siamese tracker that runs at 100 FPS on CPU and beats MixFormerV2-S by 7.6 AUC on AVisT, with evaluations across 11 benchmarks and ablations for every component. The dual-search-region, FMF layer, TRL loss, and BN-statistics TTA are new combinations of known ideas, and the tracking-specific TTA is, as far as I know, the first of its kind. Credit where due: hardware-consistent FPS measurements, external OOD benchmarks, and released code and models.\n\nThe soft spots are in the math of the inference procedure. Section 3.2 / Algorithm 2 claims a parameter-free dynamic update based on a running average of classification scores, but the printed pseudocode compares `ρt > ρt` — the score against itself — and never maintains the running average defined in Eq. (4). As printed, the dynamic update can never trigger. That is not cosmetic: Table 6 credits the update with 1.0 AUC (0.448 to 0.458), and DTTA adds another 1.4 (0.458 to 0.472), so the 47.2% headline depends on exactly the code path the paper fails to specify. Eq. (7) for DTTA also reuses symbols ambiguously, so the intended recurrence is not fully pinned down. The released code may be correct, but the paper alone is not independently reproducible for its own headline.\n\nI agree with the reader's conditional verdict. The rest of the paper holds up: the ablation story is coherent, the FMF-vs-PSA comparison is honest, and the ID/OOD trade-off is plausible. Minor concerns: no error bars or repeated runs, hyperparameters N and lambda_D are called parameter-free without sensitivity analysis, and the \"first TTA for tracking\" claim could use a more direct comparison with online-update trackers. None of these are fatal.\n\nWho is this for? People building efficient edge trackers. It is a meaningful within-subfield improvement, not a field reshape. It deserves a serious referee. I would send it to review with a request to fix Algorithm 2, clarify Eq. (7), and add sensitivity and error-bar statements. If those are addressed, the result is publishable.","headline":"A solid empirical tracking paper whose headline result depends on an underspecified test-time update that the printed Algorithm 2 cannot actually execute; worth reviewing if fixed.","tokens_in":20460,"tokens_out":2876,"would_cite":false,"duration_ms":27023,"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":"The paper claims that a ~2M-parameter Siamese tracker, SiamABC, tracks objects under out-of-distribution conditions at 100 FPS on a CPU, beating MixFormerV2-S by 7.6% AUC on AVisT while running almost 3x faster.","keywords":["visual object tracking","Siamese tracker","efficient inference","out-of-distribution generalization","test-time adaptation","batch normalization statistics","transitive relation loss","Fast Mixed Filtration"],"falsifier":"Remove DTTA while keeping FMF, the transitive relation loss, and the dual-search-region, and measure the AUC gap on AVisT and ITB. The paper's own tables show a gap of about 1.4 AUC on AVisT and 0.9 on ITB; if a reimplementation or a larger benchmark sweep finds that gap is within benchmark noise, the BN-statistic premise is not carrying the OOD claim. A more direct test is to take a tracker trained on clean weather data, evaluate on heavily corrupted variants where batch-normalization statistics are the only shifted component, and check whether the anchored instance-statistic update recovers the lost accuracy.","tokens_in":19377,"feed_emoji":"🎯","tokens_out":9624,"duration_ms":73980,"temperature":0.7,"pith_summary":"The paper claims that a lightweight Siamese tracker can track objects on out-of-distribution (OOD) video, footage unlike its training data, about as accurately as much larger models while running at 100 frames per second on a CPU. The method, SiamABC, adds a dual-search-region to keep the tracker anchored to both the initial target and its recent appearance, a learnable Fast Mixed Filtration layer to combine template and search features cheaply, a transitive relation loss for training, and a backward-free dynamic test-time adaptation that updates batch-normalization statistics during inference. On the AVisT adverse-conditions benchmark, the smallest model, S-Tiny, reaches 47.2% AUC, outperforming MixFormerV2-S by 7.6% while being almost three times faster. The paper also claims this is the first use of test-time adaptation for single-object visual tracking. If the results hold, efficient trackers could be deployed in-the-wild under constrained resources with less accuracy loss than current efficient methods.","feed_headline":"100 FPS CPU tracker tops out-of-distribution tests","feed_subtitle":"S-Tiny beats MixFormerV2-S by 7.6% AUC on AVisT at 100 FPS, 3x faster on CPU.","key_machinery":"The central object is Fast Mixed Filtration (FMF), a compact channel and spatial attention block built from 1x1 convolutions, softmax masks, and broadcast element-wise multiplications and additions. FMF filters the concatenated dual-template and dual-search-region features and compresses them from 2C channels back to C; the paper reports 0.034 GFLOPs, 0.395M parameters, and 0.4ms CPU latency, matching the accuracy of polarized self-attention at half the latency. Two other load-bearing pieces are the transitive relation loss, which pulls the filtered template and search-region representations close in cosine space using a projection head with a stop-gradient, and Dynamic Test-Time Adaptation (DTTA), a backward-free update of the head batch-normalization statistics kept anchored to the source statistics with a small anchor weight of 0.1. Together these carry the paper's accuracy and generalization claims.","core_discovery":"The central claim is that the speed-accuracy trade-off of efficient visual tracking can be pushed by bridging temporal context at both the template and the search-region ends, and by adapting the model at inference time without backpropagation. SiamABC maintains four inputs: a static template, a dynamic template, the current search region, and a dynamic search region re-centered on the object. Their features are filtered and correlated, and training is guided by a transitive relation loss. During inference, the batch-normalization statistics of the prediction heads are updated with instance statistics anchored to the source statistics. On AVisT, S-Tiny reaches 47.2% AUC at 100 FPS on a CPU, outperforming MixFormerV2-S, which gets 39.6% AUC at 37 FPS, by 7.6% while running almost three times faster.","pith_inferences":["Because DTTA only updates batch-normalization running statistics and leaves weights frozen, the same anchored-update recipe could be bolted onto other already-trained trackers at inference time, and possibly onto other single-frame video tasks, without retraining.","A controlled study that keeps the dual-search-region but removes the template update could separate the contribution of the extra temporal anchor from the contribution of simply seeing more examples.","The ablation numbers imply the OOD improvement is spread across components: FMF and the transitive relation loss add about 4 AUC points on AVisT, while DTTA adds about 1.4; a reimplementation on different corruption types would show whether the BN-statistic premise transfers beyond the tested benchmarks."],"forward_implications":["S-Tiny reaches 47.2% AUC on AVisT, ahead of MixFormerV2-S by 7.6% while running at 100 FPS instead of 37 FPS on a CPU.","The FMF layer matches polarized self-attention accuracy with about half the CPU latency (0.4ms versus 0.8ms) and lower FLOPs.","Adding the transitive relation loss raises AVisT AUC from 43.7% to 45.8% in the ablation, and the dynamic test-time adaptation lifts it further to 47.2%.","Across the 11 evaluated benchmarks the largest gains are on out-of-distribution sets such as AVisT, UAV123, ITB, and NFS30, while in-distribution performance stays competitive with efficient trackers.","DTTA is backward-free and adds negligible latency (3.7ms versus 3.6ms on a CPU), so adaptation remains compatible with real-time tracking."],"supporting_citations":[{"why":"Supplies the dual-template design and the FEAR-XS speed reference that S-Tiny matches in FPS while improving accuracy.","marker":"[6]"},{"why":"MixFormerV2-S is the efficient transformer baseline that S-Tiny surpasses by 7.6% AUC on AVisT at nearly three times the CPU speed.","marker":"[12]"},{"why":"AVisT is the adverse-visibility out-of-distribution benchmark on which the paper's main generalization claim is measured.","marker":"[45]"},{"why":"Polarized self-attention is the attention baseline that FMF matches in accuracy while reducing latency and FLOPs.","marker":"[37]"},{"why":"DUA is the backward-free BN-update baseline whose drift from source statistics motivates DTTA's anchored update.","marker":"[41]"},{"why":"AdaBN is the BN-statistics adaptation baseline that does not improve tracking in the paper's batch-size-one setting.","marker":"[33]"},{"why":"Momentum-based BN adaptation is the baseline that shows negligible improvement and motivates the anchored weighting in DTTA.","marker":"[49]"},{"why":"Instance normalization is the baseline that slows inference without an OOD accuracy gain, setting up the DTTA comparison.","marker":"[46]"}],"fun_headline_variants":["SiamABC: 100 FPS CPU tracker beats OOD baselines","Adaptive tracker generalizes to out-of-distribution video","Efficient Siamese tracker with test-time adaptation hits 100 FPS","Fast tracker improves OOD accuracy without slowing down","SiamABC: Adaptive CPU tracker for real-time OOD tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that shifts in batch-normalization statistics are a major source of out-of-distribution accuracy loss in tracking; if that premise is false for tracking, the dynamic test-time adaptation contributes little beyond the architectural and loss improvements.","fun_headline_variants_meta":{"raw":{"variants":["SiamABC: 100 FPS CPU tracker beats OOD baselines","Adaptive tracker generalizes to out-of-distribution video","Efficient Siamese tracker with test-time adaptation hits 100 FPS","Fast tracker improves OOD accuracy without slowing down","SiamABC: Adaptive CPU tracker for real-time OOD tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1472,"prompt_tokens":925,"completion_tokens":547,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":541,"completion_tokens_details":{"reasoning_tokens":459}},"tokens_in":541,"tokens_out":547,"duration_ms":6274,"temperature":1.0,"reasoning_tokens":459,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T10:49:22.842132+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Remove DTTA while keeping FMF, the transitive relation loss, and the dual-search-region, and measure the AUC gap on AVisT and ITB. The paper's own tables show a gap of about 1.4 AUC on AVisT and 0.9 on ITB; if a reimplementation or a larger benchmark sweep finds that gap is within benchmark noise, the BN-statistic premise is not carrying the OOD claim. A more direct test is to take a tracker trained on clean weather data, evaluate on heavily corrupted variants where batch-normalization statistics are the only shifted component, and check whether the anchored instance-statistic update recovers the lost accuracy.","supporting_citations":[{"cited_title":"Fear: Fast, efficient, accu- rate and robust visual tracker","cited_arxiv_id":null,"evidence_quote":"Supplies the dual-template design and the FEAR-XS speed reference that S-Tiny matches in FPS while improving accuracy."},{"cited_title":"Mixformerv2: Efficient fully transformer tracking.Advances in Neural Information Processing Systems, 36, 2024","cited_arxiv_id":null,"evidence_quote":"MixFormerV2-S is the efficient transformer baseline that S-Tiny surpasses by 7.6% AUC on AVisT at nearly three times the CPU speed."},{"cited_title":"The norm must go on: Dynamic unsuper- vised domain adaptation by normalization","cited_arxiv_id":null,"evidence_quote":"DUA is the backward-free BN-update baseline whose drift from source statistics motivates DTTA's anchored update."},{"cited_title":"Improving robustness against common corruptions by covariate shift adaptation","cited_arxiv_id":null,"evidence_quote":"Momentum-based BN adaptation is the baseline that shows negligible improvement and motivates the anchored weighting in DTTA."},{"cited_title":"Two at once: Enhancing learning and generalization capacities via ibn-net","cited_arxiv_id":null,"evidence_quote":"Instance normalization is the baseline that slows inference without an OOD accuracy gain, setting up the DTTA comparison."}],"review_version":1}