{"id":"758ed3e5-a761-457c-894d-c51d64f621bf","arxiv_id":"2505.16823","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The OWP-IMU dataset provides over 110,000 synchronized samples of infrared RSS, IMU, and ground-truth positions along continuous trajectories in LOS and NLOS indoor environments.","lead":"Researchers released an open dataset of infrared light signals, motion sensor readings, and precise indoor positions, collected by a small vehicle in two indoor environments with and without a metal obstacle. It gives indoor positioning researchers a shared testbed for comparing methods that fuse light strength and inertial measurements.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Timestamp alignment between OWP, IMU, and GT is the load-bearing assumption; the paper gives no synchronization procedure, and nearest-neighbor matching alone cannot validate it.","rationale":"The reader's weakest assumption already identified timestamp alignment and the IMU frame as the key risk. I agree that timestamp alignment is the most load-bearing concern, but I would narrow the focus: the IMU-to-PD offset of 19 mm is a constant, known quantity that can be compensated in post-processing, whereas the timestamp synchronization is unverified and could introduce time-varying errors that are much harder to correct. The paper provides no synchronization details, so the dataset's central promise of synchronized RSS, IMU, and ground-truth data is not yet established. The benchmarks and the P99 reduction claim are supporting evidence for the dataset's utility, but they would be misleading if the underlying synchronization is wrong. The concrete stationary-period test is cheap and would settle the concern without requiring access to hardware. I do not see a reason to change the reader's CONDITIONAL verdict: the concern is real but curable, and the paper's core contribution is plausible. The novelty comparison in Table I is credible given the two cited alternatives, and the self-consistency of the benchmark results (GP improves with training size; obstacles degrade accuracy) is a good sign, though not a substitute for synchronization verification.","tokens_in":7928,"tokens_out":3081,"duration_ms":27686,"concrete_test":"Use the roughly 1-minute stationary period at the start of each of the six recordings. Independently detect the stationary interval from (1) GT displacement (threshold <1 mm), (2) IMU acceleration magnitude (threshold <0.01 g), and (3) OWP RSS variance or the known LED modulation pattern. Compare the detected interval boundaries across the three streams for every trajectory. If all boundaries coincide within one GT sample (6.25 ms) and the offset does not grow over the recording, timestamp alignment is validated. If offsets appear, compute the lag and drift between streams and report whether nearest-neighbor matching introduces a position error exceeding 1 cm at the highest speed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The dataset's central value is synchronized multimodal data, yet the paper never describes how the OWP, IMU, and GT timestamp streams were synchronized. Section IV.A states only: \"for each OWP timestamp we find the closest GT timestamp and take its coordinates as the true position for that sample.\" There is no mention of a shared hardware clock, post-hoc synchronization, or drift compensation. If the clocks have even a modest relative drift, the nearest GT sample can be systematically offset, and every OWP sample would be paired with a wrong ground-truth position. Because the entire dataset and all benchmark numbers rest on this pairing, this is the most load-bearing assumption. The 19 mm IMU-to-PD offset is secondary: it is a fixed, small bias, whereas timestamp misalignment is time-varying and would contaminate every downstream use of the dataset. Notably, Figure 7's outliers could be physical (shadow or vibration) or could be artifacts of timing misalignment; without a synchronization check, the benchmark results cannot be interpreted cleanly.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents OWP-IMU, an open dataset of received-signal-strength optical wireless positioning (OWP) measurements, IMU data, and Qualisys motion-capture ground truth collected in two indoor environments (clear and with a metal-column obstacle) at three vehicle speeds. The authors describe the hardware platform, data formats, and two benchmark families: Gaussian-process regression and multilateration on RSS alone, and an extended Kalman filter fusing IMU and RSS. The central claim is that this is the first open-source dataset combining RSS-based OWP, IMU, and accurate ground truth with continuous trajectories, LOS/NLOS conditions, three velocities, over 110k samples, and more than 80 minutes of recording.","tokens_in":8118,"tokens_out":5931,"duration_ms":48206,"significance":"If the central claim holds, OWP-IMU would be a valuable community testbed: it offers high-rate OWP (27 Hz), IMU (200 Hz), and ground truth (160 Hz) with continuous trajectories and controlled LOS/NLOS conditions, and the paper includes benchmark numbers that let users anticipate performance. Strengths include a detailed mechanical platform description, use of an independent motion-capture system for ground truth, a public dataset link, and reproducible baseline methods. The main limitations are the undocumented timestamp synchronization between streams and the in-sample calibration of the multilateration benchmark, both of which must be addressed before the dataset's synchronization claim and benchmark comparisons can be fully trusted.","major_comments":[{"comment":"The paper does not describe how the OWP, IMU, and ground-truth timestamp streams are synchronized. The only statement is 'for each OWP timestamp we find the closest GT timestamp and take its coordinates as the true position for that sample.' This assumes a common clock or a known, drift-free offset between devices. If the OWP and Qualisys clocks have relative drift or a constant latency, nearest-neighbor matching will systematically pair each RSS sample to the wrong ground-truth position, and every downstream benchmark and any user of the dataset will inherit that error. Please document the synchronization method (shared clock, PTP, post-hoc alignment) and provide a validation experiment, for example a static recording showing the residual synchronization error.","section":"Section IV.A"},{"comment":"The manuscript claims 'over 110 k data points' and 'Receiver Total Captures > 110 k' (Table I), yet Section IV.A says 'The total number of synced datasets is about 17000.' At 27 Hz OWP sampling over more than 80 min one would expect roughly 130,000 OWP samples, so the two numbers cannot both describe the same set. Please clarify whether 110k is the raw OWP count, a sum over all three streams, or something else, and report per-trajectory raw and synchronized sample counts, including how the stationary initial minute and any dropped samples are treated. Without this breakdown, the dataset size and the 17k-sample benchmark cannot be interpreted.","section":"Section IV.A / Table I"},{"comment":"The multilateration benchmark is calibrated in-sample: 'We first calibrate a gain factor for each LED by all measurement datasets. Then we use it to compute each LED to PD distance from its RSS reading' and the same datasets are used for the reported errors. This gives multilateration an advantage that the Gaussian-process benchmark, which is evaluated on held-out splits, does not receive, and it makes the multilateration-vs-GP comparison and the reported multilateration errors unrepresentative of unseen conditions. Please calibrate the gain factors on a training subset or with leave-one-trajectory-out validation and report calibration and test errors separately.","section":"Section IV.A"},{"comment":"The EKF benchmark description does not state whether the 400-sample GP model used inside the EKF is trained on data from the same trajectory being evaluated or on separate training data. If the same trajectory is used for both training and evaluation, the position updates inside the EKF are trained in-sample, which can only improve the reported P50/P99 numbers. Please specify the training/evaluation split for the GP model used in the EKF and, ideally, report leave-one-trajectory-out results.","section":"Section IV.B"}],"minor_comments":[{"comment":"Table I marks reference [16] as not open access despite it being published on IEEE DataPort with a DOI; please verify and correct the table.","section":"Table I"},{"comment":"The 19 mm vertical offset between the IMU and the PD is mentioned but its effect on the EKF's position and heading estimates is not quantified; a short sensitivity note would help users.","section":"Section II / Figure 3"},{"comment":"The EKF implementation is not reproducible without the process/measurement noise covariance matrices and the GP kernel hyperparameters; please include these values or provide the code in the dataset release.","section":"Section IV.B"},{"comment":"Figure 7 attributes a large OWP error to 'shadows or vehicle vibration' without evidence; a timing-misalignment artifact could also produce such outliers, so this interpretation should be revisited once synchronization is validated.","section":"Section IV / Figure 7"},{"comment":"Minor wording: 'The total number of synced datasets is about 17000' should read 'synchronized OWP samples'; also clarify whether the GT rotation matrix is stored row-major or column-major.","section":"Section III / Table III"}],"recommendation":"major_revision","confidential_remarks":"The dataset paper is a good fit for the journal's data-descriptor-style contributions. The main risk is that the synchronization procedure is not documented; if this cannot be provided, the dataset's core claim of synchronized multimodal data is not verifiable. I also noted the comparison with prior datasets is limited to two datasets, which makes the 'first' claim weakly supported; the authors should expand the comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain take: the paper gives the field something it didn't have: an open, synchronized RSS optical wireless plus IMU dataset with continuous trajectories, orientation ground truth, three speeds, and a real NLOS obstruction. That is a useful artifact and it should become a reference for benchmarking indoor OWP fusion.\n\nWhat's new and good: Table I is honest about prior work. Compared to the 158 static points in [16] and the 98 camera AOA points in [17], this is a step change in scale and scope. The platform description is careful: LED positions, PD height, IMU-to-PD offset, motion-capture GT, stationary intervals for ZUPT. The GP and EKF benchmarks are standard, but they demonstrate the dataset is usable and give future users a baseline.\n\nSoft spots, in order. The headline \"45% reduction\" in P99 is an arithmetic overclaim: 40 cm to 25 cm is 37.5%. Fix the number. Multilateration gains are calibrated on all measurement datasets and then used to evaluate multilateration on the same data; that is in-sample and should be cross-validated. The gap between \"over 110 k\" OWP samples and \"about 17 000\" synced samples is unexplained and needs a sentence. And the benchmarks report percentiles but no variance or error bars; ten random splits should give you spread.\n\nThe timestamp synchronization question is the one I'd push hardest. The paper says only that for each OWP timestamp you find the closest GT timestamp and take that as ground truth. No shared clock, no drift compensation, no validation of the alignment. If clock drift or latency is present, every paired position is wrong in a way that is hard to see in aggregate. You don't strictly need to fix it for the dataset to be useful, but you do need to document how the streams were synchronized and provide some check, for example comparing stationary periods. The stress-test note is right that this is the load-bearing assumption, and it is currently underexplained. The 19 mm IMU-to-PD offset is much less concerning; it is small and fixed.\n\nIndoor positioning and sensor fusion researchers will want this dataset. The paper deserves serious peer review, with the benchmark cleanup and the synchronization section as required revisions. I would accept it as a dataset paper rather than desk reject.","headline":"A genuinely useful open dataset for RSS optical wireless positioning plus IMU; the benchmark claims need small but real fixes, and the timestamp synchronization must be documented.","tokens_in":8661,"tokens_out":2405,"would_cite":true,"duration_ms":18733,"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 introduces the first open-source dataset fusing optical-wireless RSS, IMU, and motion-capture ground truth for indoor positioning.","keywords":["received signal strength","optical wireless positioning","visible light positioning","indoor localization","inertial measurement unit","extended Kalman filter","Gaussian process regression","NLOS benchmark"],"falsifier":"An independent audit could take the released files and re-synchronize the streams by estimating clock offset and drift instead of taking the nearest ground-truth timestamp for each OWP sample; if the P99 errors move by more than a few centimeters, the reported 40 cm-to-25 cm improvement is partly an artifact of the alignment assumption. A second check would include the 19 mm IMU-to-photodiode lever arm in the frame alignment and recompute the EKF errors.","tokens_in":7747,"feed_emoji":"💡","tokens_out":9003,"duration_ms":63457,"temperature":0.7,"pith_summary":"This paper's objective is to give the indoor-localization community a public, synchronized dataset for received-signal-strength based optical wireless positioning combined with inertial data. The authors recorded infrared RSS at 27 Hz, IMU readings at 200 Hz, and ground-truth position and orientation at 160 Hz while a small vehicle drove continuous trajectories in two rooms, one obstacle-free and one with a metal column blocking line of sight, at three speeds. The resulting OWP-IMU dataset holds over 110,000 samples and more than 80 minutes of motion. A sympathetic reading of the paper is that these characteristics make it the first open benchmark of its kind, giving researchers a common testbed to develop and compare OWP and OWP-plus-IMU localization algorithms. The authors also show the dataset is usable by benchmarking Gaussian-process regression and extended-Kalman-filter fusion.","feed_headline":"First open dataset fuses optical RSS and IMU for positioning","feed_subtitle":"110k+ synchronized samples over 80 minutes give a shared testbed for indoor localization.","key_machinery":"The load-bearing object is the synchronized triple-stream recording: OWP RSS samples at 27 Hz, IMU readings at 200 Hz, and motion-capture ground truth at 160 Hz, collected on a vehicle with four ceiling-mounted infrared LEDs transmitting at distinct frequencies. The paper uses Gaussian-process regression to map the four RSS values to 2D position, multilateration with a Lambertian channel model as a classical baseline, and an extended Kalman filter whose state includes position, velocity, heading, accelerations, and accelerometer biases, with zero-velocity updates during the stationary segment at the start of each recording. This machinery lets the paper demonstrate both the dataset's standalone usability and the accuracy gain from fusing inertial data.","core_discovery":"The central claim is that OWP-IMU is the first open-source dataset pairing received-signal-strength optical wireless positioning with IMU data and accurate ground truth, containing continuous trajectories at three velocities under both line-of-sight and non-line-of-sight conditions. The dataset records four infrared LED RSS values at 27 Hz, IMU acceleration and gyroscope readings at 200 Hz, and motion-capture position and orientation at 160 Hz, synchronized so each OWP sample can be matched to a ground-truth position. The benchmark results show that Gaussian-process regression on the RSS alone reaches roughly 10 cm median error with 400 training samples, while an extended Kalman filter combining IMU and OWP reduces the P99 error from about 40 cm to about 25 cm and tracks heading with about 8 degrees mean absolute error over more than 10 minutes of motion.","pith_inferences":["A natural extension the authors do not pursue is using the one-minute stationary segment at the start of each recording to calibrate IMU bias per trajectory, which would test how much of the EKF gain relies on the built-in zero-velocity reset.","Because the 27 Hz RSS stream and 200 Hz IMU stream are synchronized to a common ground truth, the dataset could also support learning-based fusion methods that predict position between optical updates, a direction the paper only mentions as future work.","If the timestamp-matching assumption is violated by clock drift, the reported error numbers would shift; re-estimating the clock model from the data would reveal how much of the 40 cm to 25 cm P99 improvement is alignment-dependent."],"forward_implications":["Researchers can train and evaluate RSS-based optical wireless positioning methods on open data without building their own optical transmitter and receiver testbed.","The 27 Hz RSS and 200 Hz IMU rates with synchronized ground truth give a concrete benchmark for data-efficient learning and sensor-fusion algorithms.","GP regression reaching about 10 cm median error with 400 samples implies that small training sets suffice for fingerprint-style positioning in this environment.","The EKF results imply that adding IMU data mainly suppresses outlier errors, cutting P99 from about 40 cm to about 25 cm while keeping median error near 8 to 10 cm.","The three speeds and the obstacle versus no-obstacle contrast provide controlled conditions for studying how dynamics and non-line-of-sight conditions affect localization error."],"supporting_citations":[{"why":"An example of visible-light-positioning and inertial-integration work that motivates combined OWP-IMU methods but does not release data.","marker":"[12]"},{"why":"Another OWP-plus-inertial fusion study, cited to show that prior combined methods lack open datasets.","marker":"[15]"},{"why":"The existing open RSS-based OWP dataset, with 158 discrete points and no IMU, which OWP-IMU extends.","marker":"[16]"},{"why":"The open image-sensor OWP dataset with 98 discrete points, used to position OWP-IMU as the continuous-trajectory, IMU-inclusive option.","marker":"[17]"},{"why":"Supplies the Gaussian-process regression approach used in the RSS-only benchmark.","marker":"[18]"},{"why":"Supplies the multilateration method used as the classical baseline.","marker":"[20]"},{"why":"Supplies the stationary-update technique used in the EKF benchmark.","marker":"[25]"}],"fun_headline_variants":["Open dataset pairs optical RSS and IMU for indoor positioning","First open-source OWP-IMU dataset with 110k samples","RSS and IMU fusion benchmark dataset for indoor location","New open dataset enables RSS+IMU indoor localization testing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The benchmark numbers assume that matching each optical sample to the nearest ground-truth time stamp gives the true position and that the IMU and motion-capture frames are aligned with the 19 mm IMU-to-photodiode offset being negligible.","fun_headline_variants_meta":{"raw":{"variants":["Open dataset pairs optical RSS and IMU for indoor positioning","First open-source OWP-IMU dataset with 110k samples","RSS and IMU fusion benchmark dataset for indoor location","New open dataset enables RSS+IMU indoor localization testing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000252,"raw_usage":{"total_tokens":1539,"prompt_tokens":901,"completion_tokens":638,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":517,"completion_tokens_details":{"reasoning_tokens":568}},"tokens_in":517,"tokens_out":638,"duration_ms":5083,"temperature":1.0,"reasoning_tokens":568,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T14:54:07.816815+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"An independent audit could take the released files and re-synchronize the streams by estimating clock offset and drift instead of taking the nearest ground-truth timestamp for each OWP sample; if the P99 errors move by more than a few centimeters, the reported 40 cm-to-25 cm improvement is partly an artifact of the alignment assumption. A second check would include the 19 mm IMU-to-photodiode lever arm in the frame alignment and recompute the EKF errors.","supporting_citations":[{"cited_title":"Inertial nav igation aiding by stationary updates,","cited_arxiv_id":null,"evidence_quote":"Supplies the stationary-update technique used in the EKF benchmark."},{"cited_title":"Tightly coupled vlp/ins integrated navigation by inclination esti mation and blockage handling,","cited_arxiv_id":null,"evidence_quote":"An example of visible-light-positioning and inertial-integration work that motivates combined OWP-IMU methods but does not release data."},{"cited_title":"Towards robust visible lig ht positioning under led shortage by visual-inertial fusion,","cited_arxiv_id":null,"evidence_quote":"Another OWP-plus-inertial fusion study, cited to show that prior combined methods lack open datasets."},{"cited_title":"Machine learning for rss-based visible light positioning ,","cited_arxiv_id":null,"evidence_quote":"The existing open RSS-based OWP dataset, with 158 discrete points and no IMU, which OWP-IMU extends."},{"cited_title":"Image sensor-bas ed visible light positioning dataset,","cited_arxiv_id":null,"evidence_quote":"The open image-sensor OWP dataset with 98 discrete points, used to position OWP-IMU as the continuous-trajectory, IMU-inclusive option."},{"cited_title":"Data-efficient gaussian process regression for accurate v isible light positioning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian-process regression approach used in the RSS-only benchmark."},{"cited_title":"A survey of positioning systems using visible l ed lights,","cited_arxiv_id":null,"evidence_quote":"Supplies the multilateration method used as the classical baseline."}],"review_version":1}