{"id":"3a2eb6cb-053d-4c4a-84c0-8b83b0c786bd","arxiv_id":"2505.18490","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"DVSE estimates vehicle speed from smartphone IMU data using separate noise and pose networks trained against GNSS, and reports lower speed and distance errors than four baselines on a private crowdsourced dataset.","lead":"This paper describes a deep learning system that estimates a car's speed from a smartphone's built-in motion sensors, using GPS speed as the training signal. It reports better accuracy and generalization than earlier methods on a private dataset of 300 drivers, but the data and code are not public.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Evaluation relies on GNSS-derived speed as ground truth, yet the motivating scenarios are GNSS-degraded; the reported margin over DeepTrack may not reflect true velocity accuracy.","rationale":"The reader's weakest assumption is exactly the load-bearing concern I identify: GNSS speed is treated as both training target and evaluation ground truth without independent validation, despite the application being motivated by GNSS-degraded environments. I considered other issues, such as the absence of error bars, the teacher-forcing mismatch in the reference velocity, and the min-of-two-shifts loss matching. The missing error bars are important because the margin over DeepTrack is small (0.16 m/s at 60 s) and the 80th-percentile velocity error is essentially tied, but a multi-seed rerun would only test statistical repeatability, not whether the metric measures true velocity. The teacher-forcing and loss-matching issues are internal design choices that, while imperfect, are supported by ablations and less directly threaten the central empirical claim. The GNSS ground-truth issue is more fundamental because it underwrites every accuracy number and every comparison; if GNSS labels are biased in the target scenarios, the reported superiority could be an artifact of fitting GNSS noise. The paper's architecture is coherent, the ablations show each module contributes, and no internal contradiction is apparent, so a reject is not warranted. The appropriate outcome remains a conditional acceptance pending independent ground-truth validation, which matches the reader's original verdict; hence no change is recommended.","tokens_in":14655,"tokens_out":9073,"duration_ms":84445,"concrete_test":"Collect a validation subset of 20-50 trajectories with an independent high-rate speed reference (OBD-II/CAN bus, or RTK-GNSS plus odometer) covering urban canyons, tunnels, and open roads. Re-evaluate DVSE and DeepTrack, using the same 100-second initialization protocol for DeepTrack, against this reference on 60-second segments. Also compute GNSS-vs-reference error on the subset to quantify label noise. If the velocity MAE ranking changes, or the 0.16 m/s advantage vanishes or falls within noise, the central claim must be revised; if DVSE retains a clear lead, the GNSS ground-truth concern would be resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim in Sec. VII-D2 is that DVSE achieves the highest accuracy without initialization parameters, supported by Table I (60 s velocity MAE 2.35 vs DeepTrack's 2.51 m/s). Both the supervision target (Sec. III-A) and the evaluation metric (Sec. VII-C) are GNSS-derived speed. The paper motivates GNSS-degraded environments (tunnels, urban canyons) but does not quantify GNSS error in the 200-hour crowdsourced dataset, does not report any filtering by GNSS quality, and offers no independent reference such as OBD-II/CAN or RTK. If GNSS speed contains delays, multipath, or interpolation artifacts, the model can learn to reproduce those artifacts rather than true vehicle speed. Since the same biased signal is used for training and evaluation, the reported MAEs may reflect fidelity to GNSS rather than physical speed. This concern is amplified by the small margin: at 60 s the 80th-percentile velocity errors are 3.62 vs 3.63 m/s, nearly identical, and the 0.16 m/s mean difference is smaller than typical seed-to-seed variation in deep inertial models. Thus the 'highest accuracy' claim is not yet established for true velocity in the intended GNSS-denied settings.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes DVSE, a deep-learning framework for estimating vehicle speed from smartphone IMU data using GNSS-derived speed as supervision. The framework decomposes the task into a GRU-based noise compensation network and a TCN-based motion transformation network, adds random-rotation data augmentation to simulate arbitrary phone poses, and uses a min-of-two loss-matching scheme to tolerate timestamp misalignment between IMU and GNSS. The method is evaluated on a private 200-hour crowdsourced dataset under trajectory-split and traditional-split protocols, compared against AI-IMU, VeTorch, DeepVIP, and DeepTrack, and ablated module by module. The authors report that DVSE achieves the best accuracy without initialization, with a 60-second velocity MAE of 2.35 m/s versus 2.51 m/s for DeepTrack, and also report deployment results on Android phones via ONNX Runtime.","tokens_in":14874,"tokens_out":6747,"duration_ms":50579,"significance":"If the reported results are reliable, the paper makes a useful engineering contribution: the modular separation of pose estimation and noise compensation is well motivated by the motion model in Eqs. (1)-(5), the ablation in Table III shows that every component contributes, and the deployment measurements in Table II and Fig. 9 demonstrate a practical efficiency advantage. The random-rotation augmentation and the sliding loss-matching idea are simple and potentially transferable to other IMU-learning tasks. However, the central claims rest on a private dataset, single-run evaluations without error bars or significance tests, and GNSS-derived speed as both training target and evaluation ground truth. The margins over the strongest baseline are small, and some reported numbers are internally inconsistent. These issues must be resolved before the accuracy and generalization claims can be accepted.","major_comments":[{"comment":"The entire supervision and evaluation use GNSS-derived speed as ground truth, but the paper motivates GNSS-degraded environments (tunnels, urban canyons, underground parking) and never quantifies the GNSS error in the 200-hour crowdsourced dataset. No filtering by GNSS quality indicators (e.g., number of satellites, dilution of precision, C/N0) is reported, and no independent reference such as OBD-II/CAN, RTK, or wheel-speed data is used. As a result, the reported MAEs may reflect fidelity to GNSS artifacts (delays, multipath, interpolation) rather than true vehicle speed, especially since the 60-second advantage over DeepTrack is only 0.16 m/s and the 80th-percentile velocity errors are essentially tied (3.62 vs. 3.63 m/s). Please add a validation subset with independent ground truth, or at minimum report GNSS error statistics and quality-filtered results, and show that the main conclusions are stable under plausible GNSS noise.","section":"Sec. III-A, Sec. VII-C"},{"comment":"No error bars, confidence intervals, or significance tests are reported for any comparison. The central claim of 'highest accuracy' rests on a 0.16 m/s margin in 60-second velocity MAE, which is smaller than typical seed-to-seed variation for deep inertial models, and the 80th-percentile errors are practically identical. Please report means and variances over multiple training runs (at least five seeds) and perform paired significance tests for the comparisons in Table I, Table III, and Fig. 8. Without this, the claimed superiority over DeepTrack is not statistically established.","section":"Table I, Table III, Fig. 8"},{"comment":"The baseline evaluation protocol is not controlled enough to support the accuracy comparison. AI-IMU results are reported after 'removing the segments that converge during evaluation,' but the number and nature of removed segments are not given, which creates a risk of cherry-picking. VeTorch and DeepTrack are initialized using the first 100 seconds of each test trajectory and use GNSS to identify stationary and forward segments, so Table I does not isolate the claimed advantage 'without the need for initialization parameters'; that advantage is instead supported by Fig. 8, whose numbers do not match Table I. Please specify the exact protocol for every baseline, report how many segments are removed and why, and provide a no-initialization variant for each baseline.","section":"Sec. VII-D2, Table I"},{"comment":"The reported results are internally inconsistent. Fig. 8 reports 60-second trajectory-split DVSE errors of 2.26 m/s (velocity) and 49.75 m (distance) and DeepTrack errors of 4.15 m/s and 93.44 m, while Table I reports 60-second values of 2.35/50.84 for DVSE and 2.51/58.56 for DeepTrack under what appears to be the same trajectory-split setting. If the settings differ (e.g., with vs. without initialization, different random splits, or different input configurations), the difference must be stated explicitly; otherwise the reader cannot determine which numbers support the paper's central claim.","section":"Sec. VII-D2, Fig. 8, Table I"},{"comment":"The reference velocity input Vr creates a train/inference mismatch that is not analyzed. During training Vr is the first-frame GNSS velocity, while during inference it is the model's own previous estimate, so estimation errors can compound over the 30-60 second evaluation horizons. No experiment varies the quality of Vr or measures recursive error growth. Please add an analysis (e.g., oracle Vr versus estimated Vr, and error trajectories over longer horizons) to show that the reported accuracy does not depend on this teacher-forcing shortcut.","section":"Sec. IV-C, Sec. VI-B"}],"minor_comments":[{"comment":"There are numerous typographical and grammatical errors that should be corrected in a copyedit: 'introduct', 'drving', 'tenporal', 'varients', 'conducte', 'incorrecttion', 'donate', 'mehthod', and 'Receptive F ield' in Eq. (7).","section":"Throughout"},{"comment":"The loss weight lambda is set to 0.7 without any sensitivity analysis; please report how the results vary with lambda in a small range around this value.","section":"Sec. VI-A, Eq. (13)"},{"comment":"The data augmentation description is underspecified: the sentence 'it is important to use the same rotation matrix for each batch of data' is ambiguous, and the distribution/range of the random three-axis Euler angles (alpha_r, beta_r, gamma_r) is never given, even though it is a free hyperparameter of the method.","section":"Sec. V-C"},{"comment":"The phrase 'we remove the segments that converge during evaluation' for AI-IMU is confusing; if the intended meaning is 'segments that do not converge', it should be stated that way, along with the fraction of data removed.","section":"Sec. VII-D2"},{"comment":"The hardware is described as 'GeForce GTX 2080 Ti'; if the GPU is an RTX 2080 Ti, the name should be corrected. Also, reference [34] is cited for cosine annealing, but the cited paper is about decoupled weight decay; this reference should be replaced or corrected.","section":"Sec. VII-B2"},{"comment":"The column header 'RAMMAE' appears to be a formatting error; the RAM column should be separated from the MAE columns, and RAM units should be stated consistently (the text mentions KB and the table appears to list KB).","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"The strongest baselines, DeepTrack and VeTorch, are prior work by the same group, and the evaluation dataset is private. The contribution would be substantially more credible with a public benchmark or an evaluation on an existing inertial dataset, or with a release of anonymized data/code. I am not recommending rejection on these grounds alone, but the editor may wish to weigh reproducibility expectations for a journal-level claim of state-of-the-art accuracy."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it. The new thing is the split: a GRU noise compensation network and a TCN motion transformation network, trained with random-rotation augmentation and a min-over-alignments loss. Each ingredient is borrowed, but the combination is a real step beyond the group's earlier DeepTrack. The ablations in Table III show every module helps, and the generalization comparison on trajectory split (2.26 vs 4.15 m/s) is the most convincing evidence in the paper. The motion model is standard and the deployment results (ONNX, memory, latency) are a nice practical touch.\n\nNow the soft spots. The central claim—\"highest accuracy without initialization parameters\"—rests on a thin margin: at 60 s, velocity MAE 2.35 vs 2.51 m/s for DeepTrack, and the 80th-percentile errors are essentially tied (3.62 vs 3.63 m/s). That difference is smaller than typical seed-to-seed variation for deep inertial models, and there are no error bars or significance tests anywhere. The bigger issue is the ground truth. GNSS-derived speed is both the training target and the evaluation metric, but the paper motivates GNSS-denied environments. The authors never quantify GNSS error in their 200-hour dataset or filter by GNSS quality, so we don't know whether the model is learning true vehicle speed or reproducing GNSS artifacts. Also, the min-alignment loss: the paper doesn't specify how inference handles the alignment when no GNSS target is available—do they just use the aligned loss term? That's a missing detail. Minor: the random-rotation range is unspecified, and the AI-IMU \"remove the segments that converge\" line is ambiguous.\n\nThe stress-test concern holds up: the reported margin may not reflect true velocity accuracy in the intended GNSS-degraded setting. That said, none of this is fatal. The method is coherent, the ablations are clean, and the generalization result on trajectory split is a meaningful improvement.\n\nWho's it for? People working on phone-only inertial tracking or crowdsensing. It deserves a serious referee—the architecture and ablations are enough to warrant review, but the empirical claims need stronger evidence before publication. I'd recommend asking for error bars, a description of the augmentation distribution and inference-time alignment, and ideally a public benchmark.","headline":"Useful incremental systems paper; the architecture and ablations are sound, but the core accuracy claim over DeepTrack is held up by a small margin on a private GNSS-supervised dataset.","tokens_in":15419,"tokens_out":1980,"would_cite":false,"duration_ms":16410,"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":"DVSE claims that splitting the inertial speed-estimation problem into a noise-compensation network and a motion-transformation network, trained with random-rotation augmentation and min-alignment loss, gives the best reported…","keywords":["vehicle speed estimation","smartphone IMU","inertial sequence learning","noise compensation network","motion transformation network","random rotation augmentation","GNSS supervision","timestamp alignment loss"],"falsifier":"Collect a test drive through a tunnel or urban canyon while recording an independent wheel-speed or OBD-II reference, run DVSE on the same phone IMU, and compare. If DVSE's velocity MAE over 60 seconds is close to its reported 2.35 m/s, the claim stands; if error grows markedly relative to a GNSS comparison on the same drive, the assumption that GNSS supervision is adequate is violated.","tokens_in":14391,"feed_emoji":"🚗","tokens_out":4807,"duration_ms":31871,"temperature":0.7,"pith_summary":"The paper proposes that a smartphone can estimate a car's forward speed from its built-in IMU alone, supervised by GNSS, using a split sequence-learning architecture called DVSE. The central idea is to separate the problem into two learnable parts: a network that compensates sensor noise and bias, and a network that aligns the phone's coordinate frame with the vehicle's. On a crowdsourced dataset of about 200 hours from 300 drivers, the authors report that DVSE reaches a 60-second velocity MAE of 2.35 m/s and distance MAE of 50.84 m, improving on the previous DeepTrack baseline without needing any initialization. The claim matters because GNSS degrades in tunnels, parking structures, and dense city canyons, exactly where phone-only speed estimation is needed.","feed_headline":"Split IMU network estimates car speed without GNSS initialization","feed_subtitle":"Two lightweight neural nets beat prior smartphone-inertial trackers on a 200-hour driving dataset without initialization","key_machinery":"The split architecture is the load-bearing mechanism: a noise compensation network (a GRU with embedding and regression layers) fits the aggregate disturbance term $N$ in the forward-velocity propagation equation, and a motion transformation network (a dilated causal TCN) outputs the three Euler angles that rotate phone-frame acceleration into the vehicle frame. A random-rotation data augmentation rotates each input window by a randomly generated rotation matrix to mimic different phone poses, and a min-alignment SmoothL1 loss compares the model output with either the time-aligned GNSS target or the target shifted by one second, whichever gives the smaller error. This separation lets each network learn a simpler function, and the min-loss absorbs GNSS latency without explicit time-delay calibration.","core_discovery":"DVSE's central claim is that explicitly decomposing the inertial speed integration of Eq. (5) into a learned noise term $N$ and a learned phone-to-vehicle rotation $R_v^p$ makes supervised speed estimation both more accurate and more generalizable than learning the whole pipeline end-to-end. A GRU-based noise compensation network estimates the disturbance term from one-second feature summaries of accelerometer and gyroscope data plus a reference velocity, while a TCN-based motion transformation network estimates Euler angles from pre-integrated acceleration and gravity. Random rotation of input windows simulates arbitrary phone placements, and a loss that takes the minimum between aligned and one-second-shifted target comparisons absorbs GNSS timestamp delay. The paper reports that this combination achieves its best results without initialization parameters, and ablation experiments show each module contributes, with data augmentation giving the largest single accuracy gain.","pith_inferences":["The paper leaves implicit that the same split of pose estimation and noise compensation could transfer to other vehicle-mounted inertial tasks, such as lateral speed or attitude estimation, since Eq. (5) is specific to forward speed.","An untested but natural extension is to evaluate on GNSS-denied segments such as tunnels and parking garages against wheel-speed or OBD ground truth; the reported GNSS-supervised MAE may understate true error where GNSS itself is inaccurate.","Because data augmentation gave the largest gain, this suggests that pose diversity, not hardware diversity, is the dominant generalization bottleneck in this crowdsourced regime; whether the same holds for a wider device distribution remains open."],"forward_implications":["DVSE can be applied with no initialization or per-driver setup, unlike PCA-based or magnetometer-based baselines.","Models trained with trajectory-level splits degrade less on unseen phones and poses: DVSE's 60-second velocity MAE rose to 2.26 m/s versus DeepTrack's 4.15 m/s.","Each module is removable, and ablation shows removing data augmentation hurts most, establishing pose variation as a key generalization bottleneck.","The loss-matching mechanism tolerates GNSS delays up to about one second, reducing the need for precise timestamp filtering in training data.","The chosen GRU noise network runs on commodity smartphones with modest memory and inference time, supporting on-device use."],"supporting_citations":[{"why":"DeepTrack is the strongest baseline and prior framework; it supplies the PCA-based pose method and the generalization comparison.","marker":"[6]"},{"why":"VeTorch is a baseline that also uses TCN-based velocity estimation, and DVSE must beat it.","marker":"[4]"},{"why":"DeepVIP is a magnetometer-based baseline used in the accuracy comparison, with the LSTM variant used for replication.","marker":"[5]"},{"why":"AI-IMU is a Kalman-filter baseline requiring initialization, used to show the no-initialization advantage.","marker":"[3]"},{"why":"GRU is the core of the noise compensation network, chosen for its recurrent temporal modeling.","marker":"[23]"},{"why":"TCN is the core of the motion transformation network, chosen for dilated causal convolutions and parallel processing.","marker":"[24]"},{"why":"Spatial Transformer Network is the stated inspiration for the Motion Transformation Net's pose-output design.","marker":"[30]"},{"why":"VINS pre-integration motivates the one-second integration of accelerometer data that compresses the MTN's input.","marker":"[31]"},{"why":"The min-over-shifted-loss matching strategy is inspired by Godard's depth-estimation work and is used for timestamp mismatch.","marker":"[33]"}],"fun_headline_variants":["Split IMU net: noise and pose modules beat end-to-end","Car speed from IMU without GPS: two subnetworks win","Decompose IMU error and attitude for better speed estimates","GNSS-free speed from phone IMU via split temporal learning","Noise net plus rotation net: smarter IMU speed from phone"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that GNSS-reported speed is accurate enough to serve as both the training target and the evaluation ground truth, despite the paper's motivating scenarios being exactly those where GNSS is degraded or blocked.","fun_headline_variants_meta":{"raw":{"variants":["Split IMU net: noise and pose modules beat end-to-end","Car speed from IMU without GPS: two subnetworks win","Decompose IMU error and attitude for better speed estimates","GNSS-free speed from phone IMU via split temporal learning","Noise net plus rotation net: smarter IMU speed from phone"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000214,"raw_usage":{"total_tokens":1383,"prompt_tokens":863,"completion_tokens":520,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":479,"completion_tokens_details":{"reasoning_tokens":432}},"tokens_in":479,"tokens_out":520,"duration_ms":4998,"temperature":1.0,"reasoning_tokens":432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:30:35.767503+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Collect a test drive through a tunnel or urban canyon while recording an independent wheel-speed or OBD-II reference, run DVSE on the same phone IMU, and compare. If DVSE's velocity MAE over 60 seconds is close to its reported 2.35 m/s, the claim stands; if error grows markedly relative to a GNSS comparison on the same drive, the assumption that GNSS supervision is adequate is violated.","supporting_citations":[{"cited_title":"Vehicle inertial tracking via mobile crowdsensing: Experience and enhancement,","cited_arxiv_id":null,"evidence_quote":"DeepTrack is the strongest baseline and prior framework; it supplies the PCA-based pose method and the generalization comparison."},{"cited_title":"Glow in the dark: Smartphone inertial odometry for vehicle tracking in gps blocked environments,","cited_arxiv_id":null,"evidence_quote":"VeTorch is a baseline that also uses TCN-based velocity estimation, and DVSE must beat it."},{"cited_title":"Deepvip: Deep learning-based vehicle indoor positioning using smartphones,","cited_arxiv_id":null,"evidence_quote":"DeepVIP is a magnetometer-based baseline used in the accuracy comparison, with the LSTM variant used for replication."},{"cited_title":"Ai-imu dead-reckoning,","cited_arxiv_id":null,"evidence_quote":"AI-IMU is a Kalman-filter baseline requiring initialization, used to show the no-initialization advantage."},{"cited_title":"Spatial transformer networks,","cited_arxiv_id":null,"evidence_quote":"Spatial Transformer Network is the stated inspiration for the Motion Transformation Net's pose-output design."},{"cited_title":"Digging into self-supervised monocular depth estimation,","cited_arxiv_id":null,"evidence_quote":"The min-over-shifted-loss matching strategy is inspired by Godard's depth-estimation work and is used for timestamp mismatch."}],"review_version":1}