{"id":"cb3f2452-c3dc-43c1-af05-7844a3c9e4e8","arxiv_id":"2412.05074","paper_version":4,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A vision-based label generator turns 2D camera images into ground-truth coordinates for Wi-Fi CSI/RSSI tracking data, with a released single-antenna dataset.","lead":"This paper introduces LoFi, a camera-and-software pipeline that automatically generates position labels for Wi-Fi localization datasets from ordinary 2D images. The authors use it to release a single-antenna Wi-Fi tracking dataset with about 210,000 CSI/RSSI frames, aiming to make data-driven indoor localization cheaper and easier to collect.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The homography-based ground-plane assumption is validated only at 10 static points; dynamic motion, occlusion, or partial visibility can invalidate the bottom-center-to-feet mapping, leaving the central 'high precision' claim unsupported.","rationale":"The paper is internally consistent and the proposed pipeline is simple, low-cost, and likely useful. The code and dataset release are positive evidence, and the authors honestly report the static validation they performed. However, the central claim of 'high precision' ground truth rests on the assumption that the YOLO bottom-center pixel corresponds to the person's ground-plane position for every frame of an arbitrary walking session. This assumption is plausible for well-framed standing poses, but it is not established for the actual dataset, which contains 5 minutes of unconstrained movement per volunteer. The static test at 10 points cannot detect failures caused by dynamic gait, partial occlusion, or bounding-box truncation. The benchmark results in Table II do not close this gap because they compare predictions against the very same LoFi labels, making them a consistency check rather than an accuracy measurement. The concrete test proposed above would settle the concern in either direction: if dynamic error remains near the static error, the method's utility is confirmed; if it grows substantially, the paper's strongest claim must be revised. The reader's weakest_assumption identified the same load-bearing concern, and I see no additional objection that would justify moving beyond a conditional verdict.","tokens_in":7383,"tokens_out":3372,"duration_ms":38870,"concrete_test":"Re-run a data-collection session with the same webcam and LoFi pipeline while a volunteer walks arbitrary trajectories in the 1.8m x 4.8m region, wearing a reflective marker tracked by an independent motion-capture system (or walking along a marked tape at known speed). Compare per-frame LoFi coordinates from Eq. (3) against marker positions for at least one trajectory facing the camera, one moving away, and one with partial occlusion. If the dynamic RMSE or the 90th-percentile error exceeds roughly twice the reported static mean error (about 35 cm), the homography-based label generator is not accurate for arbitrary movement, and the paper's central claim would need to be weakened.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Equations (2)-(3) map pixel coordinates to physical coordinates via a homography built from the four rectangle corners. The physical meaning assigned to the YOLO output is the midpoint of the bounding-box bottom edge, which is treated as the person's ground-plane position. This is the load-bearing condition: if, for any frame used as ground truth, the box bottom is not the person's foot/ground contact point—due to occlusion, partial body visibility, camera angle, or walking pose—the resulting label is not the physical position. Section IV-A validates the pipeline only with two volunteers standing at 10 predefined points (mean errors 17.44, 12.13, and 11.82 cm for three views); it does not measure error during the arbitrary walking motions used to build the 210,000-frame dataset. The benchmark in Table II is trained and evaluated on LoFi-generated coordinates, so it measures self-consistency, not physical accuracy. Thus the strongest claim—that LoFi can generate high-precision ground truth—has not been demonstrated for the data the paper releases.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes LoFi, a vision-aided label generator for Wi-Fi localization and tracking datasets. A webcam image stream is processed by a YOLO object detector, the bottom-center of the detected person's bounding box is taken as the pixel-space position, and a homography built from four manually selected rectangle corners maps that point to physical coordinates. The coordinates are aligned by timestamp with CSI/RSSI frames collected by an ESP32-S3 with a single antenna. The authors release a dataset of roughly 210,000 frames from 7 volunteers moving arbitrarily in a 1.8 m by 4.8 m region, validate the coordinate conversion at 10 static points with mean errors between 11.82 cm and 17.44 cm, and report benchmark localization/classification results for CNN, ResNet, RNN, GRU, LSTM, and CSI-BERT models.","tokens_in":7605,"tokens_out":5407,"duration_ms":49916,"significance":"If the pipeline is accurate during arbitrary motion, LoFi is a meaningful practical contribution: it offers a low-cost, camera-based alternative to LiDAR or manual tagging for producing dense per-frame ground-truth coordinates for Wi-Fi sensing datasets. The released dataset and code are also valuable assets, and the demonstration that competitive localization is possible with one RX-TX pair and one antenna is interesting. The central limitation is that the paper's evidence does not yet establish the claimed high precision for the actual tracking data, and the benchmark numbers as presented are circular.","major_comments":[{"comment":"The only validation of the coordinate conversion is a static test in which two volunteers stood at 10 predefined points, and only mean errors are reported (17.44 cm, 12.13 cm, and 11.82 cm). No error bars, per-point breakdown, or number of repetitions are given, and no validation is performed during the arbitrary walking, backward walking, and standing motions that make up the released 210,000-frame dataset. Since the bottom-center of the YOLO bounding box is treated as the ground-plane position, occlusion, partial body visibility, walking pose, or camera angle can displace the label from the physical foot position; without a dynamic validation against an independent reference (for example, marker-based tracking or LiDAR), the claim that LoFi provides high-precision ground truth for the released tracking data is not supported.","section":"Section IV-A"},{"comment":"The benchmark results in Table II are circular with respect to the paper's central claim: all training and test coordinates are generated by the same LoFi pipeline, so the reported mean errors (0.58-0.94 m) measure how well each network reproduces LoFi's outputs from CSI, not how accurately those outputs represent physical positions. The static 10-point validation in Section IV-A cannot be extrapolated to these numbers. The authors should report benchmark error against an independent test set with known positions or against manually annotated held-out frames, and they should clearly state that Table II demonstrates self-consistency rather than absolute localization accuracy.","section":"Section IV-C, Table II"},{"comment":"The abstract's claim that LoFi generates ground truth 'solely from 2D images' is stronger than the method in Eqs. (2)-(3), which also requires the physical coordinates of four anchor corners, measured manually or with a ruler app. Errors in those measurements propagate to every generated label, and the paper does not describe any lens-distortion correction before applying the homography, which can bias the mapping, especially near image edges. The authors should quantify the sensitivity of the output coordinates to anchor-point measurement errors and state the distortion assumptions, or restrict the 'solely from 2D images' wording to the per-frame person detection step.","section":"Section III-A"}],"minor_comments":[{"comment":"The notation in Eq. (1) is hard to parse; please define the indicator product and the argmax over the 'person' class more explicitly.","section":"Eq. (1)"},{"comment":"The definitions of the 2, 4, and 6 classes used for classification accuracy are not given; include the region partition or a reference to it.","section":"Section IV-C"},{"comment":"The 'Scale' column mixes frames, instances, and site/antenna counts without units; please make the entries self-explanatory.","section":"Table I"},{"comment":"For reproducibility, report the camera model, the YOLO variant and weights, and whether any intrinsic or extrinsic calibration was performed before computing the homography.","section":"Section IV-A"},{"comment":"Minor language issues: 'Channel Statement Information' should be 'Channel State Information', 'LiDar' is used inconsistently, and the abstract's 'is available' should agree with the plural subject.","section":"Index terms and abstract"},{"comment":"The yellow and green star markers may be hard to distinguish in print; please use clearly distinct markers and ensure grayscale readability.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The reader's conditional verdict and the skeptic's stress test correctly identify the key weakness: the validation gap between the static 10-point test and the arbitrary-motion dataset, together with the circular benchmark evaluation. These issues are substantial but fixable within the manuscript's scope, so I recommend major revision rather than rejection. The released dataset and code are a positive contribution, and the paper's fit with the journal is acceptable if the validation concerns are addressed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper's real contribution is a released Wi-Fi tracking dataset: 210k frames, single RX-TX pair, single antenna, CSI+RSSI, free walking motions, with per-frame coordinates generated by a webcam+YOLO+homography pipeline. Second, the paper's central accuracy claim is under-supported: the homography mapping is validated only at 10 static points, and the benchmark results are self-referential.\n\nWhat's good: The integrated labeling pipeline is sensible and low-cost. The static test shows mean errors of 17.44, 12.13, and 11.82 cm across three views, which is plausible and suggests the ground-plane assumption works when people are standing and fully visible. The dataset addresses a real gap: most existing Wi-Fi localization datasets use multiple antennas or pre-defined trajectories, while LoFi offers free motion with a commodity webcam and ESP32-S3. The authors are transparent about not providing raw images for privacy, and they state code/dataset will be on GitHub.\n\nWhere it's soft: The stress-test concern lands. Equations (2)-(3) treat the bottom-center pixel of the person bounding box as the ground-plane position. That mapping is fragile under occlusion, partial visibility, camera angle, and walking posture. Validating it at 10 static points does not establish accuracy during arbitrary motion, which is exactly what the dataset contains. No error bars or per-point breakdown are given for those static errors, so even that evidence is thin. The benchmark in Table II trains and evaluates networks on LoFi-generated coordinates, so the reported sub-meter errors mostly show that networks can fit LoFi's outputs. The only non-circular evidence is the static 10-point test. A dynamic validation against an independent reference (lidar, motion capture, or even manually annotated video frames) would materially strengthen the claims.\n\nRecommendation: This deserves a serious referee. The dataset and pipeline are useful, the claims are testable, and the limitations are identifiable rather than hidden. A careful reviewer should ask for dynamic validation and a benchmark against independent ground truth. If that comes, the paper could be a solid contribution to the indoor sensing community. I'd send it to review.","headline":"A practical vision-aided labeling pipeline and a useful single-antenna Wi-Fi dataset, but the accuracy claim rests on a 10-point static test and a circular benchmark.","tokens_in":8114,"tokens_out":2385,"would_cite":true,"duration_ms":23193,"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":"LoFi generates Wi-Fi localization ground truth from ordinary 2D images.","keywords":["Wi-Fi localization","Wi-Fi tracking","channel state information","vision-aided labeling","object detection","perspective transform","indoor positioning","dataset collection"],"falsifier":"Walk a person freely through the area while simultaneously recording LoFi labels and lidar or motion-capture ground truth, then compare per-frame; if the median label error during free motion exceeds the reported static error by a wide margin or the tracking benchmark degrades when labels are corrected, the central claim fails. A simpler check: occlude the lower third of the person in some frames and observe whether the generated coordinate jumps away from the true foot position.","tokens_in":7223,"feed_emoji":"📍","tokens_out":4256,"duration_ms":38741,"temperature":0.7,"pith_summary":"This paper proposes LoFi, a label generator that derives person coordinates from 2D camera images and uses them as ground truth for training Wi-Fi localization and tracking models. The intended advantage is to replace manual tagging, predefined grids, and expensive lidar with a cheap, flexible pipeline that works with a single webcam and a commodity Wi-Fi receiver. The paper argues that if this works, data-driven Wi-Fi sensing can move beyond coarse region classification or trajectory-shape recognition toward fine-grained coordinate prediction. It supports the claim by building a dataset with an ESP32-S3 and webcam and showing benchmark networks reach mean localization errors below one meter with a single RX-TX pair and single antenna.","feed_headline":"Webcam plus YOLO generates fine-grained Wi-Fi ground truth","feed_subtitle":"LoFi turns 2D camera frames into per-frame coordinates, letting single-antenna Wi-Fi models train without lidar or hand-marked paths.","key_machinery":"The load-bearing object is the perspective transform matrix $T$ computed with OpenCV's getPerspectiveTransform from four non-collinear anchor points (Eqs. 2-3). It converts the YOLO bounding-box bottom-center pixel into a physical ground-plane coordinate. The second essential piece is timestamp-based alignment (Eq. 5) that pairs each CSI frame with the nearest image, so labels and signals share a common time base. The paper's claim rests on these two operations plus the assumption that the box bottom lies on the floor plane.","core_discovery":"The central claim is that per-frame person coordinates accurate enough to serve as ground truth can be produced solely from 2D images, with no lidar and no predefined movement paths. The system detects the person with YOLO, takes the midpoint of the bottom edge of the bounding box as the foot position in pixels, and maps it to physical space with a perspective transform computed from four hand-picked floor corners. The resulting coordinates are timestamp-aligned to CSI and RSSI frames so that each signal sample gets a position label. The paper reports label errors of roughly 12-17 cm at ten static test points and demonstrates that several standard networks can be trained on the resulting dataset to localize within about 0.6-0.9 m average error.","pith_inferences":["The paper does not directly measure label accuracy during free movement; its 10-point static check does not exercise occlusion or pose variation, so the label error during actual dataset collection is likely larger than the reported 12-17 cm.","A natural extension is to replace the single fixed homography with per-camera calibration or a depth estimate, which would relax the ground-plane assumption and help when people are partly occluded.","The same vision-labeling pipeline could be applied to multi-person scenarios by resolving person IDs across frames, turning the generator into a source for crowd-tracking datasets.","A decisive test would compare LoFi labels against lidar or motion-capture ground truth on the same arbitrary trajectories, quantifying how much of the benchmark error originates in the labels themselves."],"forward_implications":["Anyone with a webcam and commodity Wi-Fi hardware can generate dense, fine-grained labels, enabling data-driven localization without lidar or hand-marked paths.","Training data can capture arbitrary, natural movement rather than a fixed set of points or interpolated trajectories, improving model generalization.","With a single RX-TX pair and single antenna, the resulting dataset is enough to train models with mean localization error below one meter, suggesting multi-antenna setups are not always required.","The released dataset with CSI, RSSI, timestamps, coordinates, and person ID supports tracking, localization, and person identification tasks and can be sliced with arbitrary sliding windows."],"supporting_citations":[{"why":"YOLO object detector used to locate the person and produce the bounding box whose bottom-center becomes the pixel-space coordinate.","marker":"[20]"},{"why":"OpenCV's getPerspectiveTransform supplies the homography T that maps pixel coordinates to physical ground-plane coordinates.","marker":"[21]"},{"why":"Provides the packet-loss filling method and the CSI-BERT baseline; the paper's signal completion and one benchmark rely on it.","marker":"[22]"},{"why":"WiTRAJ, the prior camera-based Wi-Fi tracking work whose movement constraints LoFi is designed to avoid; appears in the dataset comparison.","marker":"[5]"}],"fun_headline_variants":["LoFi turns webcam frames into Wi-Fi training coordinates","Vision-generated labels train Wi-Fi localization without lidar","YOLO and webcam produce fine-grained Wi-Fi ground truth","2D images replace lidar for Wi-Fi localization labels","Per-frame Wi-Fi labels from a single webcam"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes the bottom-center pixel of the detected person's bounding box lies on the same ground plane as the four chosen floor corners; if feet are occluded, the body is partially out of frame, or the camera angle changes, every generated label inherits a systematic offset.","fun_headline_variants_meta":{"raw":{"variants":["LoFi turns webcam frames into Wi-Fi training coordinates","Vision-generated labels train Wi-Fi localization without lidar","YOLO and webcam produce fine-grained Wi-Fi ground truth","2D images replace lidar for Wi-Fi localization labels","Per-frame Wi-Fi labels from a single webcam"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000595,"raw_usage":{"total_tokens":2737,"prompt_tokens":850,"completion_tokens":1887,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":466,"completion_tokens_details":{"reasoning_tokens":1806}},"tokens_in":466,"tokens_out":1887,"duration_ms":12589,"temperature":1.0,"reasoning_tokens":1806,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T20:55:51.272029+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Walk a person freely through the area while simultaneously recording LoFi labels and lidar or motion-capture ground truth, then compare per-frame; if the median label error during free motion exceeds the reported static error by a wide margin or the tracking benchmark degrades when labels are corrected, the central claim fails. A simpler check: occlude the lower third of the person in some frames and observe whether the generated coordinate jumps away from the true foot position.","supporting_citations":[{"cited_title":"You only look once: Unified, real-time object detection,","cited_arxiv_id":null,"evidence_quote":"YOLO object detector used to locate the person and produce the bounding box whose bottom-center becomes the pixel-space coordinate."},{"cited_title":"Bradski, A","cited_arxiv_id":null,"evidence_quote":"OpenCV's getPerspectiveTransform supplies the homography T that maps pixel coordinates to physical ground-plane coordinates."},{"cited_title":"Witraj: Robust indoor motion tracking with wifi signals,","cited_arxiv_id":null,"evidence_quote":"WiTRAJ, the prior camera-based Wi-Fi tracking work whose movement constraints LoFi is designed to avoid; appears in the dataset comparison."}],"review_version":1}