{"id":"dab4b7ff-c429-46ad-8ccd-267e0321a4db","arxiv_id":"1908.01442","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"FOF jointly learns a Laplacian-coded compact feature representation and a dual correlation filter, improving over HCF on OTB and VOT2016 but not surpassing modern deep trackers.","lead":"This paper presents a visual tracker that learns a compact 10-channel representation of deep features jointly with a correlation filter, and tests it on three standard benchmarks. It is a useful but incremental contribution to tracking, with modest gains over older baselines and no evidence that it would change the field.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Detection/update equations (16)-(17) appear to use raw features x, not the learned code Z; if literal, the central joint-coding claim is untested at inference.","rationale":"The reader's weakest assumption, the fixed first-frame codebook, is a real limitation, but Section 4.3 argues for it and the codebook-size ablation partially probes it; it does not by itself undermine the central claim. The more fundamental issue is that, as written, the test-time procedure may not use the code Z at all. Equation (16) averages raw features x into an appearance model, and Eq. (17) forms the response from raw x and bar_x; neither references B or Z. Yet Eq. (6) learns u as the dual variable of a filter on Z. This is not a subtlety: if detection uses raw features, the compact target-oriented representation is never evaluated, and the reported improvements are not evidence for the paper's mechanism. The paper promises code and does not supply it, so a reader cannot distinguish a typographical shorthand from a genuine mismatch. The concern is concrete and testable, so the appropriate verdict remains conditional: the authors should correct or clarify the detection and update equations and make the implementation available. I therefore keep the reader's verdict but disagree with the identification of the weakest assumption.","tokens_in":13058,"tokens_out":14162,"duration_ms":140027,"concrete_test":"Take the released FOF implementation (or re-implement from Sections 4-5) and run OTB100 twice: once with detection and update exactly as in Eqs. (16)-(17) on raw x, and once with the model-consistent procedure in which every test patch is first encoded as Z_test via B (X = BZ) and the response is computed as bar_y = F^{-1}(F(bar_u) * sum_k conj(F(bar_Z^k)) * F(Z_test^k)), with the appearance model updated on Z. If the two PR/SR scores differ by more than about 1%, the paper as written does not describe the tracker that produced Table 1. If they match, Eq. (17) is a notation slip and the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that jointly learning the code Z and the filter u produces a compact target-oriented representation that improves tracking. For this to hold, the learned representation must be the one on which the filter is trained and applied. In Eq. (6), the filter u is the dual variable of a DCF defined on the encoded feature Z (via p = Zc, X = BZ). But the tracker description in Section 5.1 updates the appearance model in Eq. (16) as F(bar_x)_t = (1-eta)F(bar_x)_{t-1} + eta F(x) with x the raw image-patch feature defined in Sections 3-4.1, and computes the detection response in Eq. (17) as bar_y = F^{-1}(F(bar_u) * sum_d F(x^d * bar_x^d)). Neither step encodes the test patch against the codebook B, nor uses the coding channels Z. If these equations are literal, the learned Z plays no role at test time, so the reported PR/SR gains over HCF cannot be attributed to the proposed joint feature-coding model; they could stem entirely from the added re-detection and scale modules. Section 4.3's fixed-codebook argument becomes irrelevant in that case. Because the paper promises but does not provide code, the discrepancy cannot be resolved from the text.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes FOF, a correlation-filter tracker that learns a compact, target-oriented code Z of deep CNN features via Laplacian coding, jointly with the dual correlation filter. The central model (Eq. (6)) minimizes a DCF objective expressed on the encoded features p=Zc, plus a Laplacian regularizer on Z, subject to the reconstruction constraint X=BZ and the relation p=Zc. The optimization is carried out by ADMM; the Z-subproblem is solved by NAG, while the p- and u-subproblems have closed-form Fourier solutions. The complete tracker adds model updating, EdgeBox-based re-detection, and scale estimation to the HCF baseline, and is evaluated on OTB50, OTB100, and VOT2016. The paper reports consistent gains over HCF and state-of-the-art or comparable results, with ablation and sensitivity studies. The main technical concern, discussed below, is that the inference equations in Section 5.1 appear to operate on raw features rather than the learned code.","tokens_in":13416,"tokens_out":12921,"duration_ms":128974,"significance":"If the joint coding/filter model is genuinely the mechanism behind the reported results, this is a worthwhile contribution: it is, to the authors' knowledge, the first attempt to bring bag-of-features coding into DCF tracking, and the ADMM derivation is standard and the per-iteration cost is plausibly modest. The paper presents a fairly complete evaluation: three benchmarks, comparisons to strong baselines, an ablation study, parameter sensitivity, and a complexity analysis. The main issue is that the test-time pipeline in Section 5.1, as written, uses raw features, so the experiments do not yet establish that the reported gains come from the learned representation. Until this discrepancy is resolved, the quantitative claims cannot be interpreted as support for the proposed model.","major_comments":[{"comment":"The model update and detection are written on raw features. Eq. (16) updates the appearance model with F(x), and Eq. (17) computes the response as F^{-1}(F(\\bar u) \\odot \\sum_d F(x^d \\odot \\bar x^d)), where x is the raw image-patch feature defined in Sections 3-4.1. No step in Algorithm 1 encodes the test patch with the codebook B, and the sum in Eq. (17) runs over the D raw channels, while the dual filter \\bar u was trained on the k-channel encoded representation p=Zc in Eq. (14). As written, the learned code Z is not used at test time, so the reported PR/SR gains over HCF cannot be attributed to the joint feature-coding model; they could stem entirely from the added re-detection and scale modules. Please specify how the test features are encoded and use them in Eqs. (16)-(17), or, if the experiments actually apply the filter to raw features, state this and revise the attribution of the gains.","section":"Section 5.1, Eqs. (16)-(17), Algorithm 1"},{"comment":"The Laplacian matrix in the model is defined on the coding features Z (G_{ij}=1 if z_i is among the r nearest neighbors of z_j), but Algorithm 1 computes L_t using the raw features X_t and then solves Eq. (6) with this fixed L. The implemented objective is therefore \\gamma tr(Z L_X Z^T) rather than the stated \\gamma tr(Z L_Z Z^T). This changes the regularizer and undermines the claim of 'Laplacian coding' on the learned codes. The authors should either define the graph on the raw features and justify that choice, or compute the graph from Z during the optimization; the current description is not reproducible.","section":"Section 4.1, Eq. (6); Algorithm 1, line 5"},{"comment":"The codebook B is built from the first frame and never updated. The paper justifies this by asserting that 'the pattern across different frames would encode similar features on the fixed dictionary,' but no experiment or analysis supports this assumption. On sequences with substantial appearance change, a fixed codebook may not represent the target well, so the robustness of the learned coding is unverified. Please provide evidence (e.g., an attribute-based breakdown or tracking success as a function of appearance-change severity) or discuss conditions under which the fixed codebook fails.","section":"Section 4.3"},{"comment":"The key ablation 'noJL' is described as first performing feature coding and then training the correlation filter, but the text does not state whether this variant (and the full FOF) uses the encoded features at detection. If both variants use the raw-feature response of Eq. (17), the comparison mostly reflects differences in filter training, not joint coding at inference. In addition, the text says 'FOF over FOF-noJR' (presumably FOF-noJL), and Figure 4 is referenced without reporting its numeric values in the text. Please report the full ablation numbers and specify the inference pipeline for each variant.","section":"Section 6.5, ablation study"}],"minor_comments":[{"comment":"The text says 'cookbook' where it should say 'codebook'.","section":"Section 4.1"},{"comment":"The default \\gamma is reported as 0.8, but Table 3's sensitivity analysis varies \\gamma over 5, 10, and 15; please reconcile these values.","section":"Section 6.1 and Table 3"},{"comment":"The FCNT entries '856' and '599' are missing the leading '0.' (presumably 0.856 and 0.599).","section":"Table 1, OTB50 row"},{"comment":"The closed-form update for \\hat{p} appears to omit the 1/(MN) normalization that Parseval's theorem introduces in the \\ell_2 term; please clarify whether this factor is absorbed into \\mu or \\lambda.","section":"Section 4.2, Eq. (13)"},{"comment":"The stated O(k^3 N_J D M N) for the Z-subproblem is difficult to reconcile with the precomputation of B^T B and the NAG iterations; please provide the per-iteration cost with the actual matrix sizes.","section":"Section 4.2, complexity analysis"},{"comment":"Figure 4 is not visible in the manuscript text, so the reader cannot verify the claimed 2.9%/2.2% improvement without the figure or a table.","section":"Section 6.5, Figure 4"},{"comment":"The paper states that source codes and experimental results 'would be available online,' but this version provides no URL or release details, which limits reproducibility.","section":"Section 1 and Section 6.5"}],"recommendation":"major_revision","confidential_remarks":"The inference mismatch is severe. If the authors cannot clarify that the encoded features are used at test time, or if the experiments turn out to use raw features, the paper's central claim collapses and I would recommend rejection. I would also ask the editor to consider the inconsistency between the stated default \\gamma=0.8 and the sensitivity range 5-15, as well as the missing ablation figure. The paper claims code availability but no URL is given, which is especially important given this discrepancy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know right away: the core idea—jointly learning a Laplacian-coded feature representation and a dual correlation filter—is genuinely new, and the formulation in Eq. (6) is the first to bring BoF-style coding into the DCF framework. The optimization is standard but competently done, with an ADMM/NAG solver, convergence claims, and a complexity analysis. The empirical section is also thorough in the usual ways: three benchmarks, ablations, parameter sensitivity, and comparisons to HCF and MDNet. Credit is due for the design and the honest reporting of the extra re-detection and scale modules.\n\nBut there is a load-bearing inconsistency. The filter u in Eq. (6) is solved from the coded features Z, yet the tracker's model update (Eq. 16) and detection response (Eq. 17) both use the raw features x. There is no step that encodes the new test patch against the codebook B before correlation. If those equations are literal, then Z never appears at inference time. That means the reported PR/SR gains over HCF cannot be attributed to the proposed joint feature-coding model; they could come entirely from the re-detection and scale handling. The ablation labeled noJL tests whether training u on Z versus on a separately-coded Z changes things, but it still applies u to raw x in detection, so it does not isolate the value of using Z at test time. On top of that, Algorithm 1 computes the Laplacian matrix from the raw features X rather than from the coding features Z as Eq. (6) requires—another mismatch between model and implementation.\n\nThe paper promises source code but, as far as I can tell, none is available, so the ambiguity cannot be resolved from the text. This is a serious problem, but it is fixable: if the authors actually re-encode the patch at test time, the paper's claims become plausible and the modest gains are believable. If not, the paper is really about a filter trained on coded features but applied to raw features, which is a different and weaker claim. I would not desk-reject it, because the idea and the model are worth refereeing,\n\nand a corrected version could make a decent contribution. As it stands, though, the central claim is untested as written, and the paper needs major revision before it can be accepted. I'd still bring it to a reading group, mainly as a cautionary example of a formulation/implementation gap that reviewers need to catch.","headline":"The joint coding-plus-DCF idea is new and cleanly formulated, but the tracker as written never uses the learned code at test time, so the paper's central claim is untested by its own equations.","tokens_in":13879,"tokens_out":3278,"would_cite":false,"duration_ms":36060,"reading_group":"maybe","serious_thinker":"no","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that jointly learning a Laplacian feature code and the correlation filter produces a compact, target-oriented representation that tracks better than using raw deep features.","keywords":["visual tracking","correlation filter","feature coding","Laplacian coding","dictionary learning","ADMM","bag-of-features","target-oriented representation"],"falsifier":"Take a sequence with strong gradual appearance change, such as a target rotating and changing lighting, track with the published fixed-codebook setting, and compare against the same model but rebuilding or updating the codebook every few frames. If the fixed-codebook version loses the target significantly earlier while the updated version stays on it, the assumption fails. A cheaper check is to measure the nearest-neighbor consistency of codes for the same target patch across appearance changes and see whether similar patches produce dissimilar codes.","tokens_in":12895,"feed_emoji":"🎯","tokens_out":3983,"duration_ms":39104,"temperature":0.7,"pith_summary":"This paper tries to establish that a correlation-filter tracker becomes noticeably more robust when the raw deep features it feeds on are replaced by compact codes learned jointly with the filter, rather than computed separately. The authors argue that generic deep features are redundant and noisy for tracking, and that encoding them on a fixed dictionary with a Laplacian smoothness term yields a representation that is smaller, more stable, and target-oriented. They report that on three standard tracking benchmarks this joint scheme clearly beats the deep-feature baseline it builds on and runs at a comparable speed, and that the gain comes mostly from the joint optimization itself.","feed_headline":"Compact codes learned with the filter beat deep-feature tracking baselines","feed_subtitle":"Encoded on a fixed dictionary with a Laplacian prior, the codes stay discriminative at far lower dimension and little speed cost.","key_machinery":"The load-bearing object is the joint objective in Eq. (6): minimize over the coding matrix $Z$ and the dual filter variable $u$ the dual correlation-filter loss plus a Laplacian regularization $\\gamma\\,\\mathrm{tr}(ZLZ^\\top)$, under the reconstruction constraint $X = BZ$, with the encoded feature $Zc$ feeding the circulant matrix $S(Zc)$. The Laplacian term encodes locality and similarity among local features, and ADMM splits the problem into three convex subproblems solved by NAG and FFT. The codebook $B$ is built once by dictionary learning on the first frame.","core_discovery":"The central claim is that filter learning and feature coding should be one optimization, not two. The paper encodes each target patch as a combination of codebook atoms, feeds the code as the correlation filter's input channels, and in the same objective (Eq. (6)) fits the filter, the code, and a Laplacian prior that keeps nearby local features encoded similarly. The result, per the paper, is a feature map whose channels concentrate on different parts of the target and suppress background, so the filter becomes less sensitive to redundancy and noise in the original deep features.","pith_inferences":["The same joint-coding trick could be dropped into any circulant-shift learner, such as scale filters or re-detection filters, not just the translation filter, since the solver is generic to the dual correlation-filter form.","Because the codebook is fixed after the first frame, an online dictionary-update rule based on confident frames is a natural cheap test and may extend tracking to longer appearance drifts.","The small number of learned channels behaves like a set of part detectors; inspecting which patch regions activate each channel could give a quantitative link between tracking robustness and feature interpretability."],"forward_implications":["Replacing raw deep features with Laplacian codes of dimension $k=10$ removes most of the 512-channel noise while still improving filter discrimination, so lower-dimensional inputs need not hurt tracking.","Joint optimization contributes more than the coding alone: the ablation shows a clear drop when coding and filter are learned separately.","The unified framework keeps runtime close to the baseline, because ADMM converges in about two iterations and the added optimization is a small fraction of the frame time.","The coding step applies to multi-layer features and, the paper argues, even to low-dimensional features such as HOG or gray values, which PCA-based compression cannot improve."],"supporting_citations":[{"why":"Supplies the dual correlation-filter formulation and the FFT-based solutions that the proposed objective builds on.","marker":"[16]"},{"why":"Motivates the Laplacian coding and the stability argument that similar local features should be encoded similarly.","marker":"[14]"},{"why":"Provides the dictionary-learning algorithm used to construct the codebook from the first frame.","marker":"[19]"},{"why":"Defines the hierarchical deep-feature baseline whose features and response combination the tracker inherits.","marker":"[31]"},{"why":"Gives the ADMM framework and the multiplier-update scheme that underpin the proposed solver.","marker":"[4]"},{"why":"Serves as the state-of-the-art comparison tracker whose performance the paper matches or beats on the benchmarks.","marker":"[35]"},{"why":"Provides the OTB100 benchmark and evaluation protocol used for the main precision and success-rate numbers.","marker":"[46]"}],"fun_headline_variants":["Joint code-filter optimization trims deep features for tracking","Laplacian-coded features and filter learned in one pass","Compact target codes: filter and code fit as one","Single objective learns filter and compact target codes","One-shot filter and code learning cuts feature dimension"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a codebook built from the first frame and never updated stays good enough for the whole sequence: the paper asserts, without verifying, that appearance changes still encode similarly on this fixed dictionary, so tracking quality is not harmed.","fun_headline_variants_meta":{"raw":{"variants":["Joint code-filter optimization trims deep features for tracking","Laplacian-coded features and filter learned in one pass","Compact target codes: filter and code fit as one","Single objective learns filter and compact target codes","One-shot filter and code learning cuts feature dimension"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00025,"raw_usage":{"total_tokens":1491,"prompt_tokens":822,"completion_tokens":669,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":438,"completion_tokens_details":{"reasoning_tokens":594}},"tokens_in":438,"tokens_out":669,"duration_ms":7539,"temperature":1.0,"reasoning_tokens":594,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T15:12:53.082523+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a sequence with strong gradual appearance change, such as a target rotating and changing lighting, track with the published fixed-codebook setting, and compare against the same model but rebuilding or updating the codebook every few frames. If the fixed-codebook version loses the target significantly earlier while the updated version stays on it, the assumption fails. A cheaper check is to measure the nearest-neighbor consistency of codes for the same target patch across appearance changes and see whether similar patches produce dissimilar codes.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the dual correlation-filter formulation and the FFT-based solutions that the proposed objective builds on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the Laplacian coding and the stability argument that similar local features should be encoded similarly."},{"cited_title":"Jenatton, J","cited_arxiv_id":null,"evidence_quote":"Provides the dictionary-learning algorithm used to construct the codebook from the first frame."},{"cited_title":"Ma, J.-B","cited_arxiv_id":null,"evidence_quote":"Defines the hierarchical deep-feature baseline whose features and response combination the tracker inherits."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the ADMM framework and the multiplier-update scheme that underpin the proposed solver."},{"cited_title":"Nam and B","cited_arxiv_id":null,"evidence_quote":"Serves as the state-of-the-art comparison tracker whose performance the paper matches or beats on the benchmarks."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the OTB100 benchmark and evaluation protocol used for the main precision and success-rate numbers."}],"review_version":1}