{"id":"c267879a-541c-4ce7-91ce-79c3f3d03b2c","arxiv_id":"2508.00358","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A speed-conditioned learnable Kalman filter predicts its own noise covariances from ego-vehicle speed and object scale, improving multi-object tracking accuracy on KITTI and nuScenes.","lead":"This paper proposes a multi-object tracker whose Kalman filter noise settings are generated on the fly by a small neural network that takes the car's speed and the object's size as input, reporting top results on KITTI and nuScenes tracking benchmarks. The idea matters because a moving car's own velocity changes how much the tracker should trust its motion predictions versus its camera measurements, and speed is a free, always-available signal.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central causal role of speed is never isolated: no ablation removes the speed input while keeping MSNet, and Table 2's white-noise-speed row still gains ~4.4 HOTA, so the paper's attribution of gains to speed conditioning is unsupported.","rationale":"The reader's weakest-assumption analysis is correct and load-bearing. The paper is an empirical systems paper; the official-looking benchmark numbers and internally coherent component ablation make the overall method plausible, but the causal role of speed is the central scientific claim and it is exactly the variable that is never controlled. Table 2's noise-corrupted speed is not a valid substitute for removing speed, and Fig. 5's improvement at 0 km/h shows the mechanism operates even without a speed signal. The lack of code and the direct regression of P_t|t in Eq. 7 are secondary concerns; they affect reproducibility and the strength of the 'preserves KF structure' claim, but they are not what would overturn the headline result. I therefore keep the reader's CONDITIONAL verdict: conditional on a speed-ablation experiment and public code, but not accept or reject based on current evidence.","tokens_in":14576,"tokens_out":5467,"duration_ms":54462,"concrete_test":"Retrain SG-LKF on KITTI 2D and nuScenes with the speed channel fixed to a constant (e.g., v=0 or a learned scalar) for every frame, keeping MSNet, all three losses, track management, and hyperparameters identical, and report HOTA/AMOTA on the same test sets. If the constant-speed model matches the v*N(0,1) row (HOTA ≈ 78.8 on KITTI 2D) or comes within a small gap of 79.59, the claimed speed-driven advantage is not supported; if it drops substantially (e.g., below ~77), the speed causal claim survives this test.","verdict_should_be":"UNCHANGED","load_bearing_attack":"SG-LKF's headline claim is that conditioning P, Q, R on ego-vehicle speed is what produces the accuracy and stability gains. The paper never runs the direct control: keep MSNet and all losses but feed a constant (or no) speed signal. Table 1 ablates SG-LKF plus losses as one block against BoT-SORT, so it conflates three changes. Table 2, meant as a robustness study, is closer to a control: with v replaced by v*N(0,1), KITTI 2D HOTA is 78.84, only 0.75 below the clean-speed 79.59 and still 4.37 above the BoT-SORT baseline. A network that sees essentially random speed can still learn from object scale (w, h) and the trajectory/semantic losses, so a large fraction of the gain is not attributable to speed information. Fig. 5 compounds this: SG-LKF-3D beats SimpleTrack at 0 km/h, where the speed input carries no signal. Since Eqs. 6-7 use v, w, h as the only conditioning inputs, the sufficiency of scalar speed for ego-motion noise (ignoring bearing, depth, direction, and object motion) is untested. The authors' own limitation section concedes that >100 km/h and irregular motion are unverified. The result is that the mechanism named in the title and abstract is the least-tested part of the system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes SG-LKF, a tracking-by-detection MOT framework in which the Kalman filter's process noise Q, observation noise R, and posterior covariance P are predicted by a lightweight MLP (MSNet) from ego-vehicle speed and object box size, together with self-supervised trajectory, semantic, and positional consistency losses. The method is evaluated on KITTI 2D MOT (79.59% HOTA, first among vision-based methods), KITTI 3D MOT (82.03% HOTA), and nuScenes 3D MOT (69.00% AMOTA, +2.2% over SimpleTrack). The paper also reports speed-robustness experiments and speed-stratified comparisons against BoT-SORT and SimpleTrack.","tokens_in":14785,"tokens_out":3776,"duration_ms":38172,"significance":"The reported benchmark numbers are valuable: they come from official test servers, the component ablation in Table 1 is internally coherent, and the MSNet is extremely lightweight (78.2K parameters), making the approach practically attractive if the speed-conditioning mechanism is real. The paper also demonstrates integration into two existing KF-based trackers, which strengthens the generality claim. However, the central causal claim that ego-vehicle speed is the driver of the gains is not isolated by any experiment that removes the speed input while keeping the learnable filter and losses. Given that the title, abstract, and method definition all attribute the gains to speed conditioning, this missing control is a load-bearing gap. The manuscript is a solid systems contribution with official numbers, but the mechanism named in the title is the least-tested part of the system.","major_comments":[{"comment":"The causal role of ego speed is never isolated. There is no ablation that keeps MSNet and all losses but removes or constantizes the speed input. Table 2 is the closest to a control, and it undercuts the speed-attribution claim: with v replaced by v·N(0,1), KITTI 2D HOTA is 78.84, only 0.75 below the clean-speed value of 79.59 and still 4.37 above the BoT-SORT baseline. Since MSNet still receives (w,h) and the trajectory/semantic/position losses remain active in that condition, most of the gain is apparently not attributable to the specific speed value. The authors should add a direct control (e.g., constant speed, zero speed, or shuffled speed) with all other components unchanged and report HOTA/DetA/AssA. Without such a control, the 'speed-driven' mechanism in the title and abstract is unsupported.","section":"§4, Eqs. (6)-(7); Table 2"},{"comment":"The update stage defines P_t|t = MSNet(v_t, w_z, h_z), which replaces the standard Kalman posterior covariance update P_t|t = (I - K_t H_t) P_t|t-1 with a learned mapping. This is a structural deviation from the Kalman recursion, not merely a learned tuning of fixed parameters, and the paper provides no justification or analysis of its effect. The P/Q/R ablation mentioned in Appendix B is not included in the main text, so the reader cannot see whether the learned P_t|t is essential or even beneficial. Please either justify this design choice theoretically, or add an ablation that compares it against the standard Riccati update while keeping the same learned Q and R.","section":"§4, Eq. (7)"},{"comment":"The modeling assumption that scalar ego speed plus box scale is sufficient conditioning information for ego-motion-induced noise is untested. Physically, the effect of ego motion on a target's image-plane or BEV uncertainty depends on the angle between the ego velocity vector and the object's bearing, object depth, camera pose, and the object's own motion, none of which are inputs. The authors' own limitation section concedes that effectiveness above 100 km/h and under irregular motion is unverified. Figure 5 aggravates the concern: SG-LKF-3D improves over SimpleTrack at 0 km/h, where the speed input carries no signal, indicating that non-speed mechanisms (losses, learned P/R, or the 3D association changes) contribute substantially. The authors should test the sufficiency assumption directly, for example by holding speed constant across different bearings/depths or by including velocity direction as an additional input and showing whether gains change.","section":"§4, 'N0=[v,w,h]'; Discussion and Limitations"}],"minor_comments":[{"comment":"The checkmark columns are not fully self-explanatory. Please label the row conditions explicitly (e.g., list the active components for each row) so the reader does not have to infer which combination corresponds to 76.84, 79.49, and 79.22.","section":"Table 1"},{"comment":"The displayed equation for L_TCL is malformed in the text; the formula is split mid-sentence with symbols appearing before the 'defined as follows' clause. Please rewrite the definition cleanly.","section":"§4, Eq. (8)"},{"comment":"The phrase 'Mamba-inspired temporal aggregates' is vague: define the aggregation operation or cite the specific construction used for f_i^k and c_i^k.","section":"§4, Loss Function Design"},{"comment":"The notation 'v (1 ± 5%·N(0,1))' and 'v·N(0,1)' is ambiguous. Clarify whether the noise is multiplicative, additive, or scaled, and how each row was generated.","section":"Table 2"},{"comment":"Figure 2 is extremely dense and contains rendering artifacts (the '��' characters) in the provided version. Clean the figure and enlarge the key components so the architecture diagram is readable.","section":"Figure 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a benchmark-driven systems contribution; the official test-server numbers are a genuine asset. The main risk is over-claiming the speed mechanism: the authors need one clean control experiment that removes speed information while keeping the learned filter and losses, plus a discussion of the non-standard P_t|t update. If those are addressed, the contribution could be publishable. No code or supplementary material is provided, which limits reproducibility; the authors should be encouraged to release code and the missing Appendix B ablation."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper's actual contribution is narrower than the title. What's new: MSNet, a lightweight MLP that predicts Kalman covariances P, Q, R from ego speed and box size, plus a trajectory-consistency loss. That combination is not in the prior art (KalmanNet, HybridTrack, CMC/pose methods). The empirical work is honest: official test-server numbers on KITTI 2D/3D and nuScenes, clean integration into BoT-SORT, SimpleTrack, and VirConvTrack, and a sensible component ablation in Table 1. MSNet is 78.2K parameters, so efficiency is real.\n\nThe soft spot is the central claim. There is no ablation that removes the speed input while keeping MSNet and the losses. Table 2 is the closest control, and it undercuts the speed narrative: replacing v with v·N(0,1), effectively white noise, still gives 78.84 HOTA — only 0.75 below the clean-speed run and 4.37 above BoT-SORT. So most of the gain can be explained by object scale and the losses, with speed contributing little. Fig. 5 reinforces this: SG-LKF beats SimpleTrack at 0 km/h, where speed input is constant. The model's assumption that scalar v plus w,h is sufficient for ego-motion noise — ignoring bearing, depth, direction, and object motion — is untested. Also, Eq. 7 predicts the posterior covariance directly rather than computing it through the Kalman gain; that's a design choice that needs justification. The margins over prior SOTA are tiny (79.59 vs 79.53 HOTA on KITTI 2D). The limitations section candidly admits >100 km/h and irregular motion are unverified.\n\nFor whom: MOT researchers working on adaptive Kalman filtering, and people integrating lightweight modules into existing trackers. It's a plausible engineering contribution, not a foundational one.\n\nRecommendation: send to peer review with a request for the missing control — a learned filter with speed removed or constant — and an analysis of what the network learns from speed. If speed survives that, the paper becomes solid. As is, the title overclaims.","headline":"Speed-conditioned Kalman filtering is a new idea, but the paper never isolates speed from scale and losses, so the headline mechanism is unproven even though the full system works.","tokens_in":15457,"tokens_out":2798,"would_cite":false,"duration_ms":28090,"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":"This paper claims that adapting a Kalman filter's noise covariances to the ego vehicle's own speed stabilizes multi-object tracking and sets the top reported HOTA among vision-based methods on KITTI 2D MOT.","keywords":["multi-object tracking","Kalman filter","ego-vehicle speed","learnable noise covariance","tracking-by-detection","autonomous driving","KITTI benchmark","nuScenes benchmark"],"falsifier":"Train SG-LKF with the speed input $v$ replaced by a constant while keeping MSNet and all three losses identical and rerun the KITTI and nuScenes benchmarks; if HOTA and AMOTA do not drop materially, the reported gains are not caused by speed conditioning. A second check is to split sequences by the angle between the ego-velocity vector and the target's bearing at equal speeds; if the gain disappears when the vehicle moves across the target's line of sight rather than toward it, the scalar-speed premise is incomplete.","tokens_in":14252,"feed_emoji":"🚗","tokens_out":12211,"duration_ms":103654,"temperature":0.7,"pith_summary":"Autonomous vehicles need to keep identifying the same object across frames even while the car itself is moving, and standard trackers often lose objects or swap identities at speed because their motion model assumes a fixed reference frame. This paper's central claim is that the ego vehicle's speed is a usable signal: feeding current speed and the tracked object's box size into a small network that sets the Kalman filter's noise covariances keeps the filter's uncertainty honest as speed changes. On KITTI 2D MOT the resulting tracker reports 79.59% HOTA, the top result among vision-based methods, and the same speed-conditioning idea lifts 3D trackers on KITTI and nuScenes. If the claim holds, existing Kalman-filter trackers can be made more stable on moving vehicles at almost no extra cost, since vehicle speed is already available on board.","feed_headline":"Car speed tunes Kalman noise to top KITTI 2D MOT","feed_subtitle":"Conditioning tracking noise on ego speed and object size cuts ID switches and stays stable from 0 to 60 km/h.","key_machinery":"The central object is the Speed-Guided Learnable Kalman Filter (SG-LKF), a standard Kalman filter whose covariance parameters are produced on the fly by MotionScaleNet (MSNet), a decoupled token-mixing and channel-mixing MLP with inputs $N_0=[v, w, h]$ (plus object length $l$ in the 3D version). MSNet outputs positive semi-definite values for $P$, $Q$, and $R$, so the filter reacts to speed rather than running with fixed noise. The second mechanism is a self-supervised trajectory consistency loss, combined with a semantic consistency loss computed from DINOv2 box embeddings and a position consistency loss based on CIoU, which together enforce temporal smoothness, appearance coherence, and spatial alignment during training; none of these require a re-identification network at inference.","core_discovery":"The discovery is that the three covariance matrices that control a Kalman filter for multi-object tracking — the predicted-state covariance $P$, the process noise $Q$, and the observation noise $R$ — can be generated on each frame by a tiny MLP, MotionScaleNet (MSNet), whose only inputs are ego-vehicle speed $v$ and the tracked object's width and height (plus length in 3D). With those covariances adapted per frame, the filter widens its uncertainty when ego speed would inflate prediction error and tightens it when the object is stable, and the paper shows higher HOTA, DetA, and AssA at every ego speed from 0 to 60 km/h compared with BoT-SORT. The same module transplanted into VirConvTrack and SimpleTrack raises their 3D tracking scores, which the paper takes as evidence that speed-conditioned covariance generation is a general mechanism rather than a dataset-specific trick.","pith_inferences":["Beyond the paper: the scalar-speed premise suggests a concrete refinement — conditioning on the ego velocity vector (speed plus heading) and the target's bearing or depth should capture the geometry of ego-motion noise more fully; this is testable by comparing gains when the vehicle drives toward versus across a target.","Beyond the paper: if the mechanism is the cause of the gains, then any tracker that already outputs object scale can adopt the same noise-parameterization pattern, making the contribution a reusable design rather than a dataset-specific fix.","Beyond the paper: the reported experiments do not isolate the speed input from the new losses, so an ablation with speed held constant while keeping MSNet and the trajectory loss would reveal how much of the 2.37 HOTA gain is specifically due to speed conditioning.","Beyond the paper: the authors' limitation statement means the 'stable at any speed' claim is only verified up to about 60 km/h on KITTI and 35 km/h on nuScenes; highway-speed data above 100 km/h is the natural stress test."],"forward_implications":["Tracking-by-detection systems can absorb ego-vehicle speed without extra sensors, since speed is already available on the vehicle and the learnable covariance module adds only 78.2K parameters.","Speed conditioning should flatten the performance-versus-speed curve: the paper's comparison shows BoT-SORT's HOTA and AssA falling sharply as ego speed rises, while SG-LKF's decline is much slower.","The mechanism transfers across modalities: the same parameter-generation idea works for 2D image-plane tracking and for 3D LiDAR-based tracking on KITTI and nuScenes.","Because the Kalman equations are unchanged, the module can be inserted into other KF-based trackers; the reported 2.2% AMOTA gain over SimpleTrack on nuScenes is a direct demonstration.","The training-only use of DINOv2 semantic features means the association gains do not require running a re-identification network at inference time."],"supporting_citations":[{"why":"Defines the SORT tracking-by-detection paradigm whose fixed Kalman filter is the starting point that SG-LKF generalizes.","marker":"Bewley et al. 2016"},{"why":"Provides the BoT-SORT baseline and the KITTI detections; SG-LKF-2D's state definition and speed analysis are built on it.","marker":"Aharon, Orfaig, and Bobrovsky 2022"},{"why":"KalmanNet is the main learned-Kalman contrast: it learns gains or states directly, while SG-LKF instead learns P, Q, and R and keeps the Kalman equations intact.","marker":"Revach et al. 2022"},{"why":"SimpleTrack supplies the 3D baseline on nuScenes whose geometry-aware association SG-LKF-3D inherits, with a 2.2% AMOTA gain.","marker":"Pang, Li, and Wang 2022"},{"why":"VirConvTrack is the KITTI 3D baseline that gains 0.47 HOTA when SG-LKF is integrated, supporting cross-method transfer.","marker":"Wu et al. 2023"},{"why":"DINOv2 features feed the semantic and trajectory consistency losses during training, which the ablation links to higher AssA.","marker":"Oquab et al. 2023"},{"why":"CIoU defines the position consistency loss that aligns predicted and ground-truth boxes in the joint objective.","marker":"Zheng et al. 2021"}],"fun_headline_variants":["Speed-Based Kalman Noise Boosts Tracking Accuracy","Ego Speed Tunes Kalman Covariance for Better MOT","Learnable Kalman Filter Adapts to Vehicle Speed","Speed-Conditioned Tracking Hits Top KITTI 2D Score","Kalman Filter Noise Scaled by Speed Improves MOT"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method assumes that one scalar number — the vehicle's own speed, combined with the tracked object's box size — is enough to describe the extra uncertainty that the vehicle's motion injects into the tracker, so the learned adjustments transfer to any speed, direction, and driving geometry.","fun_headline_variants_meta":{"raw":{"variants":["Speed-Based Kalman Noise Boosts Tracking Accuracy","Ego Speed Tunes Kalman Covariance for Better MOT","Learnable Kalman Filter Adapts to Vehicle Speed","Speed-Conditioned Tracking Hits Top KITTI 2D Score","Kalman Filter Noise Scaled by Speed Improves MOT"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000248,"raw_usage":{"total_tokens":1567,"prompt_tokens":987,"completion_tokens":580,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":603,"completion_tokens_details":{"reasoning_tokens":497}},"tokens_in":603,"tokens_out":580,"duration_ms":5908,"temperature":1.0,"reasoning_tokens":497,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T10:13:01.426823+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train SG-LKF with the speed input $v$ replaced by a constant while keeping MSNet and all three losses identical and rerun the KITTI and nuScenes benchmarks; if HOTA and AMOTA do not drop materially, the reported gains are not caused by speed conditioning. A second check is to split sequences by the angle between the ego-velocity vector and the target's bearing at equal speeds; if the gain disappears when the vehicle moves across the target's line of sight rather than toward it, the scalar-speed premise is incomplete.","supporting_citations":[],"review_version":1}