{"id":"1d937ad5-6b81-4c2a-80cb-80f05657d250","arxiv_id":"1908.02231","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"ARCF extends the BACF correlation filter with a temporal penalty on response map changes, improving UAV tracking accuracy on three benchmarks.","lead":"This paper adds a penalty term to a standard UAV tracking algorithm that forces the tracker's confidence map to change smoothly between frames, which suppresses sudden failures during occlusion and background clutter. On three aerial tracking benchmarks with 243 sequences, the new tracker beats the baseline and several published trackers while running fast enough for real-time use.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. 12's Sherman-Morrison update contradicts Eq. 11; the coefficient gamma* = mu/(1+gamma)^2 is wrong, so the stated ADMM solution is not the minimizer of Eq. 5 unless the code silently differs.","rationale":"The reader's weakest assumption about the previous response map being corrupted is legitimate and is acknowledged by the paper's own Fig. 8, where lost-object aberrances persist. However, the more load-bearing weakness is the internal inconsistency between Eq. 11 and Eq. 12: without a correct ADMM update, the paper's stated optimization is not actually solved by the described algorithm. This is concrete, checkable, and sits at the center of the claimed contribution, whereas the response-map-reference issue affects only hard failure modes. I therefore keep the reader's conditional verdict, but anchor it on a different condition: verification of the released code against the corrected Sherman-Morrison formula. The availability of code is real credit, but it does not remove the need to settle whether the manuscript's central equation is what the implementation runs.","tokens_in":13553,"tokens_out":10069,"duration_ms":109984,"concrete_test":"Re-derive Eq. 12 symbolically from Eq. 11 using Sherman-Morrison, then evaluate both formulas on the scalar probe D=1, x=y=M=1, zeta=0, w=0, mu=1, gamma=0.71; the outputs should match if the simplification is correct, and they do not. Next, inspect the released MATLAB code at https://github.com/vision4robotics/ARCF-tracker and record the exact coefficient used in the g-hat update (search for the Sherman-Morrison branch). If the code matches Eq. 12, rerun ARCF-H on a subset of UAV123 and compare with the paper's reported numbers; if it matches the corrected coefficient, treat Eq. 12 as a typo and request an erratum. As a secondary check, recompute Table 2's map difference from ground-truth-aligned response maps rather than from the tracker's own outputs, so that the repression metric is not identical to the training penalty.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim depends on the ADMM closed-form update solving Eq. 5. Equation 11 is a correct solution of the per-frequency subproblem in Eq. 10, but Eq. 12, presented as its Sherman-Morrison simplification, has the wrong coefficients. Applying Sherman-Morrison to Eq. 11 with A = mu/(1+gamma) I and u = v = x gives g = (1/mu) R - [(1+gamma)/(mu(mu+(1+gamma)Sx))] x (x^T R), where R = x y + gamma x M - zeta + mu w and Sx = x^T x. Equation 12 instead uses gamma* = mu/(1+gamma)^2 and divides the second term by b = Sx + mu/(1+gamma), which is not equivalent. A scalar probe (D=1, x=y=M=1, zeta=0, w=0, mu=1, gamma=0.71) yields 0.631 from Eq. 11 but 0.216 from Eq. 12. If the released MATLAB code literally implements Eq. 12, the trained filter is not the minimizer of Eq. 5 and the reported gains are not explained by the stated optimization; if the code uses a corrected formula, Eq. 12 is a serious reproducibility error that still requires an erratum. A secondary issue is that Table 2 measures the exact quantity being penalized in the loss, so the lower map difference for ARCF is partly by construction rather than independent evidence of aberrance suppression.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Starting from the BACF tracker, the paper augments the DCF learning objective with a temporal regularization term that penalizes differences between the previous frame's peak-aligned response map and the current response map (Eqs. 2–3). It solves the resulting convex problem with ADMM in the Fourier domain (Eqs. 5–13), updates the appearance model online (Eq. 14), and evaluates two feature configurations (HOG-only and HOG+CN+gray) on UAV123@10fps, DTB70, and UAVDT against 20 existing trackers. The reported results show consistent precision and success-rate gains over BACF at real-time CPU speeds, and the authors interpret these gains as evidence that the response-map change penalty suppresses 'aberrant' detection peaks caused by occlusion, background clutter, and appearance changes.","tokens_in":13896,"tokens_out":9145,"duration_ms":86938,"significance":"The proposed modification is simple and well motivated, and if the implementation is consistent with the stated objective, the consistent improvements over BACF on three UAV benchmarks, together with the CPU-only speed measurements, would be a useful contribution to real-time UAV tracking. The paper's strengths include making the code public, reporting speed, and evaluating on 243 sequences, including attribute-based analyses. However, the ADMM derivation contains a concrete algebraic error, and the response-map-difference table measures the very quantity the loss minimizes; these issues must be addressed before the central claim is fully supported.","major_comments":[{"comment":"Eq. (12) is not algebraically equivalent to Eq. (11), which is the correct solution of the per-frequency subproblem. With A = μ/(1+γ) I_D and u = v = x̂(n), the Sherman–Morrison formula yields (1/(1+γ))(A + x̂ x̂^T)^{-1} = (1/μ) I - [(1+γ)/(μ(μ+(1+γ)Sx))] x̂ x̂^T, where Sx = x̂^T x̂. The coefficients in Eq. (12) are instead γ* = μ/(1+γ)^2 for the first term and γ*/b with b = Sx + μ/(1+γ) for the second, which do not match. For example, with D=1, x=y=M=1, ζ=0, w=0, μ=1, γ=0.71, Eq. (11) gives 0.631 while Eq. (12) gives 0.216. The stated ADMM update therefore does not minimize the objective in Eq. (5); please correct Eq. (12), check the released code against the corrected formula, and state which update the experiments used.","section":"§4.3.2, Eq. (12)"},{"comment":"The average map difference in Table 2 is computed with Eq. (2), the exact quantity added to the BACF objective in the third term of Eq. (3). A lower value for ARCF-H is therefore partly by construction and does not provide independent evidence that aberrance repression is what improves tracking. The precision/success gains on UAV123@10fps, DTB70, and UAVDT are independent of this term and do support the mechanism; please reframe Table 2 as a consistency check of the optimized objective rather than as a validation of the central claim.","section":"§5.4, Table 2"},{"comment":"The abstract states that ARCF 'has proven itself to have outperformed other 20 state-of-the-art trackers based on DCF and deep-based frameworks,' but the deep-tracker comparison is presented only for UAVDT (Fig. 7); on UAV123@10fps and DTB70, ARCF is compared only with hand-crafted trackers. Please either add the deep-tracker comparisons on all three datasets or qualify the claim so that it matches the reported evidence.","section":"Abstract and §5.3"}],"minor_comments":[{"comment":"The regularizer assumes the previous frame's response map, after peak alignment, is a reliable reference; this assumption can fail under full occlusion or out-of-view, as Fig. 8 itself shows. Please add an explicit limitation sentence in the formulation section and, ideally, a confidence-gating mechanism.","section":"§4.1, Eq. (3)"},{"comment":"The paper reports CPU-only speeds for hand-crafted trackers, but it is unclear whether the deep trackers in Fig. 7 were run under the same protocol; please clarify because otherwise the real-time comparison is not apples-to-apples.","section":"§5.3 and Table 1"},{"comment":"The manuscript contains numerous typographical issues, including 'trainig', 'aberrrance', and 'UA V' with an unintended space in the title and throughout; a careful proofreading pass is needed.","section":"Throughout"},{"comment":"The public code link is given without a version or commit identifier; please include a tagged release to improve reproducibility.","section":"Section 5.1"}],"recommendation":"major_revision","confidential_remarks":"To the editor: the algebra in Eq. (12) is wrong, but this is repairable by an erratum and a code check; I therefore recommend major revision, not rejection. If the released code turns out to implement the erroneous update, the authors should demonstrate that the empirical results are unchanged with the corrected formula. The circularity of Table 2 is a presentation issue that can be fixed by reframing."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague — the ARCF paper is worth your attention for one idea: regularizing the temporal change in the response map, with a peak-alignment shift, inside a DCF training loop. That is genuinely new. LMCF and ACFN use response maps to decide when to update or switch, not to constrain the learned filter. Applying this regularizer on top of BACF is a natural extension, and the empirical gains over BACF on three UAV benchmarks look credible. The experiments are broad and the paper is clearly written.\n\nThe soft spot is not the idea; it is the optimization. The stress-test note is right: Eq. 11 correctly solves the per-frequency ADMM subproblem, but Eq. 12, offered as its Sherman-Morrison simplification, has the wrong coefficients. Using the printed γ* = μ/(1+γ)^2 and b = Sx + μ/(1+γ) does not reproduce Eq. 11. A scalar probe gives 0.631 from Eq. 11 versus 0.216 from Eq. 12. That is not a cosmetic typo. If the released code literally implements Eq. 12, the trained filter is not the minimizer of the stated objective, and the reported gains are not explained by the paper's loss. If the code uses a corrected formula, then Eq. 12 is a serious reproducibility error that still demands an erratum. I could not run the code, so I cannot tell which case holds. This needs to be checked before any citation of the optimization.\n\nTwo smaller issues. Table 2 reports the exact quantity the loss minimizes, so lower \"map difference\" for ARCF is partly by construction; the independent evidence is the benchmark accuracy, which supports the mechanism, but the table itself should not be over-read. Second, the abstract's claim of beating all 20 trackers including deep ones is only supported by UAVDT; on UAV123 and DTB70 the headline comparison is against hand-crafted trackers. That overclaim should be toned down. Also, there is no sensitivity analysis for γ, η, or the ADMM iteration count.\n\nWho is this for? DCF tracker researchers and UAV tracking practitioners. It deserves a serious referee: the idea is useful, the evaluation is substantial, and the Eq. 12 problem is concrete enough that a competent reviewer can resolve it. I would recommend sending to peer review, but require the authors to reconcile Eq. 12 with the released code and the stated objective before acceptance.","headline":"The response-map regularizer is a genuine and useful idea, but the ADMM derivation has a concrete error in Eq. 12 that must be checked against the code before the optimization claims can be trusted.","tokens_in":14420,"tokens_out":4475,"would_cite":false,"duration_ms":40578,"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":"Aberrance-repressed correlation filters keep UAV trackers from jumping to background clutter.","keywords":["visual object tracking","UAV tracking","discriminative correlation filter","aberrance repression","response map regularization","background-aware correlation filter","real-time tracking","occlusion handling"],"falsifier":"Run ARCF and BACF on the full-occlusion or out-of-view subsets of UAV123 and measure center error in the first ten frames after the target reappears; if ARCF reacquires more slowly or drifts in more sequences than BACF right after occlusion, the previous-response-map assumption is the failure point.","tokens_in":13371,"feed_emoji":"🚁","tokens_out":8859,"duration_ms":87149,"temperature":0.7,"pith_summary":"This paper tries to establish that a correlation-filter tracker for drones can be made more robust by directly penalizing sudden changes in its response maps. The proposed ARCF tracker adds a regularization term to the BACF objective that compares the current response map with the previous one, after aligning their peaks, and suppresses the abrupt jumps that accompany occlusion, background clutter, and appearance change. If the claim holds, aerial trackers can handle challenging situations without relying on deep networks, at CPU speed suitable for real-time use. The paper supports it with experiments on 243 sequences containing over 90,000 frames from three UAV benchmarks, reporting favorable precision and success rates against 20 comparison trackers.","feed_headline":"Smoother response maps keep drone trackers locked on target","feed_subtitle":"ARCF adds a penalty on abrupt response-map jumps, beating 20 trackers on UAV benchmarks in real time.","key_machinery":"The load-bearing device is the aberrance penalty, defined as the squared Euclidean norm of the difference between the previous frame's response map and the current one, with a peak-shift operator $[\\psi_{p,q}]$ that slides the previous map so its peak coincides with the current peak. Added to the BACF objective in Eq. 3 with strength $\\gamma = 0.71$, this term is transformed into the frequency domain and minimized by ADMM with closed-form updates for the two subproblems, accelerated by the Sherman-Morrison formula. The shift operator makes the penalty translation-invariant, so it punishes changes in the response map's shape and confidence rather than the target's motion.","core_discovery":"The paper's central claim is that the response map produced during detection carries information earlier DCF methods discard: when an abrupt change or wrong peak appears, the filter is about to learn from a bad location. ARCF adds to the BACF objective the penalty $\\frac{\\gamma}{2}\\|\\mathbf{M}_{k-1}[\\psi_{p,q}] - \\sum_d \\mathbf{B}\\mathbf{x}_k^d \\star \\mathbf{w}_k^d\\|_2^2$, which forces the current response map to stay close to the previous one after the two peaks are aligned. Because this repression is part of the training objective and is solved by ADMM, it changes the filter that is learned rather than merely post-processing detections. The paper reports that this lowers the average response-map difference by about 20%, 24%, and 15% on UAV123@10fps, DTB70, and UAVDT, and that ARCF-HC ranks first among the tested trackers on those datasets.","pith_inferences":["An untested extension is to make $\\gamma$ adaptive: lower it when the tracker is confident so genuine fast motion is not smoothed away, and raise it when confidence drops during suspected occlusion.","The peak-aligned shift makes the penalty invariant to translation, so the same construction could be extended to penalize response-map shape changes under scale or rotation by aligning maps more richly before comparison.","If the regularization is as effective as reported, it should also stabilize deep-feature DCF trackers, potentially letting drone trackers use richer features without sacrificing real-time speed."],"forward_implications":["Because the aberrance penalty is a plug-in regularization term, the same idea should transfer to other DCF trackers; the paper explicitly names ECO and SRDCF as possible recipients.","On the occlusion attribute, ARCF-HC improves AUC over BACF by 8.1% on UAV123@10fps, 9.8% on DTB70, and 5.2% on UAVDT, so the benefit is concentrated where response maps jump the most.","The average response-map difference measured by Eq. 2 falls by roughly 20%, 24%, and 15% on the three datasets, giving a direct quantitative readout of aberrance repression independent of final tracking accuracy.","ARCF-H runs at 51.2 FPS on CPU, so the added penalty costs little speed and the method remains deployable on UAV hardware."],"supporting_citations":[{"why":"The BACF baseline whose cropping-matrix objective and ADMM optimization ARCF inherits and extends with the aberrance-regularization term.","marker":"[15]"},{"why":"Provides the UAV123@10fps benchmark and its evaluation protocol, which supplies the primary dataset for the reported comparisons.","marker":"[20]"},{"why":"Provides the DTB70 benchmark, the second dataset used to evaluate the tracker and its occlusion attributes.","marker":"[17]"},{"why":"Provides the UAVDT benchmark, the third dataset used for evaluation and attribute-based analysis.","marker":"[10]"}],"fun_headline_variants":["Penalty on response-map jumps boosts drone tracking","ARCF represses response-map aberrances for robust UAV tracking","Restrict response-map change to keep drone trackers on target","Response-map smoothness penalty makes drone tracking more reliable"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The penalty assumes the previous frame's response map is still a trustworthy picture of the target; if that map was corrupted by full occlusion, a wrong detection, or the target leaving the view, the penalty pulls the current filter toward the corrupted map and can slow recovery.","fun_headline_variants_meta":{"raw":{"variants":["Penalty on response-map jumps boosts drone tracking","ARCF represses response-map aberrances for robust UAV tracking","Restrict response-map change to keep drone trackers on target","Response-map smoothness penalty makes drone tracking more reliable"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000718,"raw_usage":{"total_tokens":3251,"prompt_tokens":996,"completion_tokens":2255,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":612,"completion_tokens_details":{"reasoning_tokens":2189}},"tokens_in":612,"tokens_out":2255,"duration_ms":17187,"temperature":1.0,"reasoning_tokens":2189,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T14:50:33.052097+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run ARCF and BACF on the full-occlusion or out-of-view subsets of UAV123 and measure center error in the first ten frames after the target reappears; if ARCF reacquires more slowly or drifts in more sequences than BACF right after occlusion, the previous-response-map assumption is the failure point.","supporting_citations":[{"cited_title":"Learning background-aware correlation filters for visual tracking","cited_arxiv_id":null,"evidence_quote":"The BACF baseline whose cropping-matrix objective and ADMM optimization ARCF inherits and extends with the aberrance-regularization term."},{"cited_title":"A benchmark and simulator for uav tracking","cited_arxiv_id":null,"evidence_quote":"Provides the UAV123@10fps benchmark and its evaluation protocol, which supplies the primary dataset for the reported comparisons."},{"cited_title":"Visual object tracking for un- manned aerial vehicles: A benchmark and new motion mod- els","cited_arxiv_id":null,"evidence_quote":"Provides the DTB70 benchmark, the second dataset used to evaluate the tracker and its occlusion attributes."},{"cited_title":"The unmanned aerial vehicle benchmark: object detection and tracking","cited_arxiv_id":null,"evidence_quote":"Provides the UAVDT benchmark, the third dataset used for evaluation and attribute-based analysis."}],"review_version":1}