{"id":"b6924980-f9ed-43c4-aa49-892e737ae089","arxiv_id":"2607.25146","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":4,"one_line_summary":"FIDAC merges multi-model face detection, human coding, and plane benchmarking to extract interpersonal distance from ordinary 2D video.","lead":"FIDAC is an open-source pipeline that turns ordinary 2D video into interpersonal-distance estimates by combining face detectors, optional human coding, and a simple depth benchmark. It aims to give social-science researchers a low-cost way to track proxemics without VR headsets or marker systems.","discovery_kind":"new_method","skeptic_critique":{"model":"moonshotai/kimi-k3","headline":"The §2.3 benchmarking tool reduces 2D-to-metric conversion to a single two-point scale ratio, which is only exact for fronto-parallel motion at the calibration location — depth and lateral displacement introduce uncorrected, unquantified error.","rationale":"The reader identified exactly this soft spot: sufficiency of a single known-distance pair on the participants' plane for accurate metric conversion under unknown depth, pose, and camera geometry. My analysis sharpens it technically (perspective-induced scale variation, face-vs-floor plane offset, and the fact that IPD-relevant motion is depth motion), but it is the same load-bearing assumption, so I agree with the reader. I do not recommend moving the verdict from CONDITIONAL: the paper is a short tool announcement that explicitly schedules accuracy evaluation (\"evaluating FIDAC's performance at various depths and orientations\") as future work, so it does not overclaim validated accuracy in the text — the risk is that users of the \"easy-to-use\" tool will treat outputs as metrically accurate before that evaluation exists. The CONDITIONAL verdict (acceptable as a tool announcement contingent on code release and added evaluation) already prices this in. Two secondary issues reinforce rather than change the verdict: no repository link despite \"open-sourced\" language, and no stated convention for how face bounding-box centers are mapped onto the calibrated plane. If the proposed depth-variation test shows large errors, the benchmarking section would need redesign (homography + camera-height handling), at which point a stronger verdict would be warranted; if errors are small, the tool's core claim is substantially supported at near-zero cost.","tokens_in":5322,"tokens_out":1251,"duration_ms":37742,"concrete_test":"Film two seated/standing actors at tape-measured ground-truth distances (e.g., 0.5/1.2/2.5/4 m, spanning Hall's zones) with a smartphone at a typical oblique tripod angle, calibrating FIDAC once on a table edge as in Figure 3. Repeat with actors displaced 0.5–1 m toward and away from the camera and laterally across the frame. If FIDAC's reported IPD error exceeds ~10% or crosses a Hall-zone boundary at off-calibration depths, the single-ratio calibration is insufficient and a ground-plane homography (4+ points) must replace it before metric claims are usable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The pipeline's entire claim to output \"actionable data about location and interpersonal distance\" rests on §2.3: the user pauses the video, clicks two points \"on the same plane as the participants\" whose real distance is known, and FIDAC applies that as a ratio against pixel distances. Three problems are unaddressed. (1) Under perspective projection, pixels-per-meter is not constant across a plane — it varies with image position and depth. A single global scale factor is exact only for fronto-parallel planes at the calibration depth; participants seated across a table will rarely satisfy this with a handheld or tripod smartphone shot at an oblique angle. (2) The measured quantity is the separation of face bounding boxes, which sit above the floor plane; unless the camera is at face height and faces stay in one plane, projecting face positions onto the calibrated plane introduces additional error the paper never discusses. (3) IPD dynamics of interest (one person leaning in, approaching) are precisely motions that change depth relative to the camera, moving the subject off the calibration depth — so the measurement degrades exactly where the signal lives. The correct minimal fix (a four-point homography on the ground plane, plus a stated assumption about projecting faces to feet) is standard and cheap, but as described the tool's metric accuracy is unknown and the authors defer all evaluation to future work. The paper is honest about this, so this is a correctness-risk concern, not an inconsistency.","agreement_with_reader":"agree"},"referee_report":{"model":"moonshotai/kimi-k3","summary":"The manuscript presents FIDAC, an open-source Python library that extracts interpersonal distance (IPD) from ordinary 2D video. The pipeline has three stages: frame extraction via OpenCV, face detection using pre-loaded models (Py-Feat and MediaPipe, with user-swappable alternatives) whose outputs are combined, and a merge step in which model-generated detections are supplemented by a human-coding interface for frames where detection fails. A \"benchmarking\" tool (§2.3) converts pixel distances to metric units by having the user click two points of known real-world separation on the participants' plane. The paper is a methods description: no accuracy, precision, or error characterization is reported, and the abstract and Conclusion explicitly defer evaluation (varied depths/orientations, comparison against existing detectors, a user study) to future work.","tokens_in":5673,"tokens_out":2124,"duration_ms":109276,"significance":"If the tool performs as described, it addresses a real and well-motivated gap: proxemics researchers currently choose between expensive motion capture, VR telemetry, or tedious manual coding, and an accessible open-source pipeline built on commodity video would lower the barrier substantially. The manuscript ships working software (open-sourced), a sensible human-in-the-loop design for detection failures, and is commendably honest about its own limitations — the authors state plainly that validation at varying depths and orientations is future work rather than overselling the current release. However, the contribution as it stands is a system description without quantitative evidence that the output distances are accurate, and the one component that converts pixels to meters (§2.3) rests on a geometric assumption that is generally violated in realistic filming conditions. The significance is therefore conditional: the engineering is plausible and useful, but the central promise of \"actionable data about location and interpersonal distance\" is not yet demonstrated.","major_comments":[{"comment":"The metric conversion reduces to a single two-point scale ratio (known distance / pixel distance). Under perspective projection, pixels-per-meter is not constant across a plane: it varies with image position and with depth. A single global scale factor is exact only for a fronto-parallel plane at the calibration depth. In the paper's own motivating example (two participants seated across a table filmed by a smartphone), an oblique camera angle makes the near and far ends of the table subtend different pixels-per-meter, so the same physical distance yields different pixel distances depending on where in the frame the participants sit. The authors should either (a) implement the standard minimal fix — a four-point planar homography on the ground/table plane, which is cheap and available in OpenCV — or (b) state the fronto-parallel assumption explicitly and provide an error bound or worked","section":"§2.3 Benchmarking, Figure 3"},{"comment":"The quantity measured is the separation of face bounding boxes, which sit above the calibrated plane (e.g., above the tabletop or floor). Unless the camera is at face height and faces remain in a single plane parallel to the image plane, projecting face positions onto the calibrated plane introduces an additional, unmodeled error. This matters most exactly where the signal lives: the IPD dynamics of interest (leaning in, approaching) change the subject's depth relative to the camera, moving them off the calibration depth and inflating or deflating apparent distance. The manuscript never discusses this geometry. At minimum the paper needs a stated model of how 2D face-box positions are interpreted as participant positions (box centers? feet estimated from box bottom?) and a discussion of when the approximation breaks down.","section":"§2.3 / §2.2 interaction"},{"comment":"Two load-bearing algorithmic details are unspecified. (1) Identity association: IPD requires knowing which bounding box belongs to which participant in every frame, but no tracking or data-association method is described — the text only says faces are detected per frame. For two or more people who move, cross, or temporarily occlude each other, naive per-frame detection will swap identities and corrupt the distance time series. (2) Multi-model fusion: the text says models are combined so that 'discrepancies in one are corrected by another,' but no fusion rule is given (union? intersection? confidence-weighted vote? temporal gap-filling?). Since the multi-model merge is one of the paper's three headline contributions, the merge logic needs to be specified precisely enough to reproduce.","section":"§2.2 Analysis + Facial Detection"},{"comment":"No quantitative evaluation of any kind is reported — no detection accuracy on a benchmark clip, no distance error versus ground truth at any depth or orientation, not even a worked example on a staged video with a tape-measure reference. The authors are transparent that evaluation is future work, which is appropriate for a demo or workshop track, but for a journal publication the central claim ('transforms facial detection results into actionable data about location and interpersonal distance') requires at least minimal evidence. A low-cost experiment is within the paper's scope: film two people at known separations (e.g., 0.5–4 m) at a few camera angles, and report measured-vs-true distance error. This would also directly address the §2.3 concerns above.","section":"Whole manuscript; Abstract and §3"}],"minor_comments":[{"comment":"Hall's zone boundaries are given as exact centimeter cutoffs (e.g., intimate 0–46 cm); since FIDAC's output will be compared against these zones, a sentence acknowledging that zone boundaries are approximate cultural averages would help readers interpret threshold-adjacent measurements.","section":"§1"},{"comment":"The phrase 'FIDAC's algorithms were trained to identify faces' is misleading — FIDAC wraps pre-trained detectors (Py-Feat, MediaPipe); it does not train them. Please reword to avoid implying the authors trained the detection models.","section":"§2.1"},{"comment":"Guidance on the free parameters (confidence threshold, min/max face size, frames-per-second) is qualitative only. Default values used by the authors, and a note on how frame subsampling interacts with the temporal resolution needed for IPD dynamics, would aid reproducibility.","section":"§2.2"},{"comment":"'Unitary values' should presumably read 'real-world units' or 'metric values.'","section":"§3"},{"comment":"The manuscript states the library is open-sourced but no repository URL, license, or version appears in the text provided; please add a link (anonymized if required). An OpenCV citation is also missing. Figures 1–3 are described but their resolution/legibility should be checked at print size.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"This reads as a competent early-stage tool paper from a group with a strong track record in proxemics/VR research. The work is honest about its unvalidated status, which I count in its favor, but in its current form it is closer to a software demo or workshop paper than a journal article: the single-scale-factor metric conversion in §2.3 is the load-bearing weakness, and both the homography fix and a minimal ground-truth validation are cheap and within scope. If the venue has a systems/demo track, that may be a better fit; otherwise I would expect a revision with at least one quantitative experiment before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"Punchline: this is a short methods note packaging known face detectors into a researcher-facing IPD pipeline. It is useful as a workflow description, not as evidence that the distances are accurate.\n\nWhat is new is the integration, not the detectors. They wire OpenCV frame sampling, MediaPipe and Py-Feat (with optional multi-model merge, confidence and size filters), a human-coding path for misses, and a two-point plane scale into one named library aimed at proxemics people who do not want markers or VR headsets. The writing is clear, the motivation from Hall’s zones and stop-distance limits is fair, and the citations are appropriate. They are also upfront that evaluation is future work—no overclaiming of results they do not have.\n\nThe soft spots are real and central, not nitpicks. There are no accuracy, precision, or error-vs-depth/orientation numbers at all. The load-bearing metric step is a single known distance between two clicked points used as a global pixels-to-meters ratio. That is exact only under strong assumptions (roughly fronto-parallel motion at the calibration depth). Face boxes sit above the floor plane; leaning and approach change depth exactly where IPD signal lives; perspective scale is not constant across the image. A four-point homography plus a stated face-to-plane projection would be the minimal standard fix, and they do not do it. “Open-sourced” is asserted without a repo link in the manuscript, which weakens reproducibility right now.\n\nWho it is for: nonverbal-behavior and HCI labs that already shoot table-top or room video and want a starting pipeline plus human fallback. Not for anyone who needs calibrated metric IPD out of the box today.\n\nI would send it to peer review as a tool/methods short paper, with the expectation that referees demand code release and at least a basic depth/orientation error study before acceptance. Engage if you care about measurement tooling in proxemics; skip if you need validated geometry.","headline":"Honest tool announcement for IPD-from-video, but metric claims rest on an unvalidated single-scale benchmark and zero accuracy numbers.","tokens_in":6596,"tokens_out":509,"would_cite":false,"duration_ms":15089,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"FIDAC turns ordinary 2D video into usable interpersonal-distance data by merging facial detectors, optional human coding, and a plane scale.","keywords":["nonverbal behaviors","facial detection","video-based tracking","proxemics","interpersonal distance","human coding","2D video analysis"],"falsifier":"Run the full pipeline on videos where true interpersonal distances are independently measured at several depths, angles, and partial occlusions, and check whether the recovered distances stay within a stated error band against those ground-truth values.","tokens_in":6437,"feed_emoji":"👥","tokens_out":803,"duration_ms":26858,"temperature":0.7,"pith_summary":"Interpersonal distance carries a lot of social meaning, but ordinary camera video does not hand it over as clean numbers. This paper presents FIDAC, an open pipeline that takes everyday 2D video, runs multiple facial detectors, fills gaps with a human-coding path when faces are missed, and converts pixel separations into real-world distances with a simple two-point benchmark on the participants’ plane. The aim is to give researchers who study nonverbal behavior a practical way to track who stands where over time without specialized motion-capture hardware. If the method works as described, dynamic proximity and related cues become extractable from smartphone or digital-camera recordings that labs already collect. The authors position the tool as a foundation they will later test across depths and orientations and extend toward broader proxemic measures such as synchrony.","feed_headline":"Open pipeline turns phone video into social distance data","feed_subtitle":"Merge face detectors, optional human coding, and a plane scale to track who stands where over time.","key_machinery":"FIDAC’s three-step pipeline: frame extraction, multi-model facial detection with confidence and size filters plus optional human coding of misses, then merge and plane benchmarking that converts bounding-box pixel separations into real-world units.","core_discovery":"The paper claims that combining several open-source facial detection models, a structured human-coding fallback for missed faces, and a benchmarking step that scales pixel distances from two known points on the same plane is enough to turn ordinary 2D video into actionable location and interpersonal-distance time series.","pith_inferences":["If the planned depth-and-orientation benchmark holds, archival interview and meeting video could be reanalyzed for proximity dynamics that were never instrumented at capture time.","The human-coding fallback implies a hybrid accuracy model: automated bulk tracking plus sparse expert correction may be more practical for many social-science labs than fully automatic 3D systems.","Failure modes at extreme angles or heavy occlusion will likely set the real ceiling on when FIDAC can replace dedicated motion capture rather than complement it."],"forward_implications":["Labs can extract continuous interpersonal-distance traces from ordinary smartphone or digital-camera recordings instead of only stop-distance snapshots.","Missed or blocked faces need not drop entire segments if the human-coding path is used to complete the series.","Pixel face locations become interpretable in Hall-style proximity zones once the plane benchmark is applied.","The same position series can later feed automated proxemic features such as synchrony once those modules are added.","Researchers can swap in other facial detectors without rebuilding the rest of the distance pipeline."],"fun_headline_variants":["FIDAC merges face detectors to pull interpersonal distance from video","Open library turns 2D video into location and social-distance time series","Combine face models, human coding, and plane scale for distance data","Pipeline extracts who-stands-where tracks from ordinary phone video","FIDAC: face detection fusion plus coding yields actionable proxemic data"],"cache_read_input_tokens":128,"weakest_assumption_plain":"A single known real-world distance between two user-picked points on the same plane as the people is enough to turn 2D face-box pixel gaps into accurate interpersonal distances despite unknown depth, pose, and camera geometry.","fun_headline_variants_meta":{"raw":{"variants":["FIDAC merges face detectors to pull interpersonal distance from video","Open library turns 2D video into location and social-distance time series","Combine face models, human coding, and plane scale for distance data","Pipeline extracts who-stands-where tracks from ordinary phone video","FIDAC: face detection fusion plus coding yields actionable proxemic data"]},"model":"grok-4.5","effort":"low","cost_usd":0.004082,"raw_usage":{"total_tokens":1160,"prompt_tokens":665,"num_sources_used":0,"completion_tokens":92,"cost_in_usd_ticks":40824000,"prompt_tokens_details":{"text_tokens":665,"audio_tokens":0,"image_tokens":0,"cached_tokens":128},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":403,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":665,"tokens_out":92,"duration_ms":7422,"temperature":1.0,"reasoning_tokens":403,"cache_read_input_tokens":128,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-31T00:07:26.692945+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Run the full pipeline on videos where true interpersonal distances are independently measured at several depths, angles, and partial occlusions, and check whether the recovered distances stay within a stated error band against those ground-truth values.","supporting_citations":[],"review_version":1}