{"id":"94346da7-f4b2-4912-a5ca-6f5329798e15","arxiv_id":"1909.02211","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"The paper derives a pixel-to-meter scale from the known gravitational acceleration measured on a jumping person's center of mass, achieving a 3.9 cm median error on on-spot jumps without camera calibration.","lead":"A person's height can be measured from ordinary video by watching them jump: gravity's known acceleration is used as a ruler to convert pixels to meters. This removes the need for camera calibration or reference objects, but only works when the person is captured in free fall from a static camera.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The scale q inferred from a jump is applied to h_px measured in standing frames without verifying that person-camera depth is unchanged; Table 3's lateral-motion errors show this omitted precondition is load-bearing.","rationale":"The reader's weakest_assumption identifies exactly the same load-bearing concern: the scale q is measured from the jump while h_px is measured from the standing frames, and the conversion is only valid if the person-camera depth is unchanged and the camera is static. The paper does state the static camera assumption, but it never lists the depth-constancy requirement, and the abstract's broad 'jumping motions' phrasing overstates the scope. The theoretical derivation itself is internally consistent under scaled orthographic projection, so this is not a fatal flaw; it is a boundary condition that the paper partially documents through its lateral-motion results in Table 3. The reader's CONDITIONAL verdict already accounts for this overstatement and for the unverified depth assumption, so my analysis does not change the verdict. I agree with the reader rather than disagreeing: the depth-constancy issue is the single most load-bearing condition on the central claim, and the proposed test would settle whether it is indeed the dominant error source in realistic use.","tokens_in":13195,"tokens_out":6083,"duration_ms":67222,"concrete_test":"Take one subject from ArticulatedFreeFall with a visible reference object. Compute q_jump from the on-spot jump's vertical COM acceleration and h_px from the first standing frames. Independently measure the true scale q_standing and q_jump from the reference object's pixel height at the standing location and at the jump location. If the method's height estimate h_px * q_jump deviates from ground truth by more than (q_jump - q_standing) * h_px, the depth-constancy assumption is the main error source. To make the test decisive, record an additional sequence where the subject steps 1 m closer to the camera between the standing frames and the jump; the gravity method should then show a bias of approximately (1 - d_after/d_before) * true_height, confirming the omitted precondition.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim requires that the pixel-to-meter scale q = g/a_px obtained from the free-fall parabola at the jump location is identical to the scale needed to convert the standing-frame pixel height h_px to metric height. Under the paper's scaled-orthographic model, q = d/f for an upright camera, so this holds only if the person's distance d from the camera is the same during the standing measurement and during the jump. The paper never states this as a precondition; Section 4.4 simply says h_px is measured 'at the first frames of the video, assuming an upright stance.' On-spot jumps in the dataset keep d approximately fixed, which is why Table 1 achieves 3.9 cm MAE. But Table 3 shows the failure mode: lateral running and jumping, where d changes, give MAE of 12.3–18.9 cm even though the free-fall scale is still measured during flight. The abstract's claim of 'MAE of up to 3.9 cm on jumping motions' therefore applies only to the depth-constant subset, not to jumping motions in general. This is not an internal inconsistency in the derivation; it is an unstated boundary condition on the central claim. If a user applies the method to a video where the subject walks toward the camera and then jumps, the estimated height will be systematically wrong by roughly the ratio of the two depths.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes to estimate a person's metric height from monocular video by using the gravitational acceleration g as an absolute reference. Under a scaled-orthographic projection model, the pixel acceleration measured during a free-fall phase yields a conversion factor q = g / a_px, which is then applied to the pixel height of a standing person to recover metric height h = c * h_px * q. The method is validated on a falling tennis ball and on a new dataset of 12 subjects performing on-spot jumps, lateral runs, and lateral jumps at two distances, reporting a median MAE of 3.9 cm on on-spot jumps. The paper also demonstrates the use of the height estimate to rescale a monocular 3D pose estimator and makes code and data available.","tokens_in":13316,"tokens_out":6489,"duration_ms":66322,"significance":"The core idea is elegant and the derivation (Eqs. 1-7) is clean and internally consistent under scaled orthographic projection. A notable strength is that the scale factor q requires no camera intrinsics, extrinsics, ground-plane knowledge, or any reference object in the scene, making it a genuinely complementary cue to appearance-based or geometry-based height estimation. The appendix simulation, the ablation study, and the public release of the ArticulatedFreeFall dataset are valuable contributions. If the boundary conditions of the method are stated precisely and the headline claims are qualified, the paper would offer a useful new tool for forensic, surveillance, and monocular reconstruction applications.","major_comments":[{"comment":"The derivation assumes a single constant scale q = d/f, but the algorithm estimates q from the free-fall segment and applies it to h_px measured 'at the first frames of the video' (Section 4.4). This is valid only if the person's distance to the camera is identical at the two times. The paper never states this constant-depth precondition; the limitation section names the static-camera assumption but not the depth-constancy requirement. The appendix simulation (Table 4) shows the magnitude of the resulting failure: for a jump straight toward the camera at d = 4 m, the error is 21 cm even in a noise-free simulation. This is a load-bearing boundary condition on the central claim, and it should be stated explicitly. The authors should either restrict the claim to on-spot/fronto-parallel motions or add an experiment in which the subject changes depth between the standing measurement and the jump, demonstrating the expected degradation.","section":"Section 3.1, Eq. (7), and Section 4.4"},{"comment":"The abstract and Section 5.2 state 'reaching a MAE of up to 3.9 cm on jumping motions.' According to Table 1, this number is the median over four on-spot jumps per subject. Table 2 shows that per-jump MAE is substantially higher: for example, low jumps have MAE of 13.1 cm at 4 m and 20.7 cm at 7 m, and even the best single-jump category remains above 5 cm. The headline therefore overstates the accuracy available from a single jump, which is the realistic operating condition for many applications. The abstract should be reworded to say 'a median over four on-spot jumps per subject yields 3.9 cm MAE,' and the per-jump variability should be highlighted more prominently in the discussion.","section":"Abstract and Section 5.2 (Tables 1 and 2)"}],"minor_comments":[{"comment":"There is a typo in 'demonstrates the success our our method'; 'our our' should be 'of our'.","section":"Section 1"},{"comment":"The sentence 'The naive Baseline-population-mean, predicting the population mean height of 168.9 m' should read '168.9 cm'.","section":"Section 5.2"},{"comment":"Table references are inconsistent: the accuracy and ablation results are discussed as 'Table 4', but the relevant tables in the main text are Tables 1 and 3; the appendix also contains a Table 4, which is confusing.","section":"Sections 5.2 and 8"},{"comment":"The word 'dominante' should be 'dominates'.","section":"Section 3.1"},{"comment":"The sentence 'the COM estimation generally decreases with the dirstance' contains a typo: 'dirstance' should be 'distance'.","section":"Section 8"},{"comment":"The caption line 'Distance d in [m] 400' appears malformed; the axis label and numeric tick should be formatted as a normal axis.","section":"Figure 7 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper's central idea is sound and the dataset release is a positive contribution. The main risk is that the abstract and conclusion overstate the operational accuracy and omit a key boundary condition. I believe these issues are fixable within the scope of a revision, so I do not recommend rejection."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read on 1909.02211.\n\nThe thing to know: this paper actually delivers something new — a way to turn free-fall acceleration into a pixel-to-meter scale for uncalibrated monocular video, applied to articulated humans via COM tracking. The derivation in Sec. 3 is simple and internally consistent under scaled orthographic projection. The authors are honest that it only recovers height, the gravity-aligned extent, and not arbitrary 3D scale. The new 12-subject dataset and the ball-dropping sanity check are useful contributions. Credit where due: the COM argument is well made, the RANSAC/polynomial fitting for acceleration is sensible, and the paper beats the appearance-based baseline [13] on their own data with lower bias.\n\nSoft spots, in order of importance. First, the abstract's 'MAE of up to 3.9 cm' applies to a median across four on-spot jumps per subject, not to per-jump predictions. Table 2 shows low jumps have MAE of 13–21 cm. That's not hiding a broken method, but it's an optimistic summary. Second — and this is the one I'd want fixed before publication — the scale q is measured during the jump, while the height h_px is measured in the first standing frames. That only works if the person-camera distance is the same in both. The paper never states this as a precondition. The authors mention 'static camera' and 'free fall', but not constant depth between the two measurements. The lateral-motion results in Table 3 (12–19 cm MAE) are consistent with depth change being a load-bearing failure mode, even though the authors attribute them to occlusion and short flight phases. The appendix simulation covers non-fronto-parallel trajectories, but doesn't test the exact standing-vs-jump depth mismatch. Third, the comparison with [13] uses different aggregation protocols for the two methods (median over jumps vs median over all frames), so the 'nearly half the error' claim should be taken with a grain of salt. Fourth, flight-phase detection requires a hand-selected snippet; the paper says this, so it's a stated limitation, not a hidden one.\n\nNone of these kill the central idea. For on-spot jumps with roughly constant depth, the method works, the math checks out, and the errors are reasonable. People working on monocular metric scale estimation, forensic video analysis, or person re-id will get value from this paper. It deserves a serious referee. With revisions — state the constant-depth condition explicitly, redo the abstract numbers, and align the comparison protocol — it would be a solid contribution. I'd send it out.","headline":"A clever and sound physics-based scale estimator for monocular video, but the headline accuracy is a median over selected on-spot jumps and the constant-depth precondition is understated.","tokens_in":14027,"tokens_out":3654,"would_cite":true,"duration_ms":31115,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The known acceleration of gravity can convert pixel measurements in ordinary video into a person's metric height, without calibrating the camera or ground plane.","keywords":["height estimation","gravity reference","free-fall","monocular video","scale recovery","center of mass","camera uncalibrated","jumping motion"],"falsifier":"A controlled video of a person of known height jumping at constant distance to a static camera should reproduce the paper's 3.9 cm MAE; the same person jumping while walking one meter toward the camera during the clip should produce a systematically biased estimate if the distance-consistency assumption breaks.","tokens_in":12828,"feed_emoji":"📏","tokens_out":7311,"duration_ms":67009,"temperature":0.7,"pith_summary":"This paper claims that the omnipresent constant $g = 9.81\\,\\mathrm{m/s^2}$ can replace a calibration object when measuring a person's height from video. It shows that the quadratic term of a projected free-fall trajectory, measured purely in pixels, can be equated to $g$ to yield a scale factor $q$, and that multiplying a standing person's pixel height by $q$ recovers metric height with a mean absolute error of 3.9 cm on jumping motions. The method requires only that the video contains a free-fall phase, that the camera is static, and that air friction is negligible. Its appeal is that it works on legacy, uncalibrated monocular footage where existing methods need extra sensors, reference objects, or manual calibration.","feed_headline":"Video of a jump reveals height to 3.9 cm without calibration","feed_subtitle":"Earth's constant gravitational acceleration turns pixel motion into meters on any uncalibrated video.","key_machinery":"The load-bearing identity is the ratio between the known gravitational acceleration and the measured image acceleration: $q = g / a_{px}$, obtained by matching the quadratic coefficient of the projected free-fall parabola (Eq. 4 and Eq. 5). The companion identity $h = h_{px} q$ (Eq. 7) then rescales any pixel measurement aligned with gravity into meters. The argument is carried through for people by replacing the body with its center of mass, whose projected trajectory must still follow the same parabola regardless of internal articulation.","core_discovery":"Under scaled-orthographic projection, a rigid object in free fall traces a parabola on the image plane. Writing the projected trajectory as $p_{px}(t)=\\frac{1}{2}a_{px}t^2 + v_{0,px}t + p_{0,px}$, the paper equates the measured image acceleration $a_{px}$ to the known gravitational constant $g$ through $q = g/a_{px}$ (Eq. 5). Because any pixel length aligned with gravity equals metric length times $1/q$, the standing person's head-to-heel pixel extent $h_{px}$ converts to height $h = h_{px} q$ (Eq. 7). This requires no knowledge of camera intrinsics, object distance, or gravity direction. For articulated people, the same derivation is carried out on the center of mass, using conservation of momentum so that internal body motions do not invalidate the projectile model.","pith_inferences":["The same ratio-symmetric identity would apply to any scene containing a moving object whose only external acceleration is known (e.g., a dropped tool), not just to human jumpers.","A practical self-check suggested by this reasoning: if the person moves toward or away from the camera between the jump and the standing measurement, the estimate degrades; errors of 12.3–18.9 cm in the running experiments indicate that distance consistency can serve as a quality control metric.","One could extend the method to temporally varying scale by estimating $q$ continuously during the jump and rescaling each frame's height, rather than transferring a single $q$ to standing frames, which would avoid the distance-consistency assumption.","For applications like person re-identification, the unbiased, per-person height estimate could serve as a geometric fingerprint complementary to appearance cues."],"forward_implications":["Metric height becomes recoverable from uncalibrated monocular video of a jump, without reference objects or camera/ground-plane calibration.","Legacy video and consumer footage can be rescaled to meters, making forensic or surveillance height estimates possible on footage that existing calibrated methods cannot process.","The recovered scale factor $q$ can be applied to the output of an off-the-shelf monocular 3D pose estimator to turn it into metric 3D pose, which the paper demonstrates in a qualitative experiment.","Because $q$ is measured per jump, taking the median over several jumps reduces the effect of occasional pose-detection failures and improves accuracy."],"supporting_citations":[{"why":"Supplies the free-fall equation of motion that yields the quadratic trajectory and the known constant $g$.","marker":"[22]"},{"why":"Provides the per-frame 2D keypoint detections used to compute the center-of-mass trajectory.","marker":"[8]"},{"why":"Provides the body-segment mass ratios used to compute the center of mass from keypoints.","marker":"[4]"},{"why":"RANSAC outlier rejection stabilizes the quadratic curve fit of the trajectory.","marker":"[9]"},{"why":"The appearance-based single-image height regression baseline whose 6.5 cm MAE the paper compares against.","marker":"[13]"}],"fun_headline_variants":["Jump video measures height within 3.9 cm, no calibration","Gravity as ruler: video jump gives height estimate","No-cal height from jump: gravity does the work","Video jump uses earth's gravity to gauge person's height","Height from video: gravity replaces calibration"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The scale ratio measured during free fall is applied to the standing height measured in a different set of frames; this only works when the person's distance to the camera is the same in both parts of the video and the camera does not move.","fun_headline_variants_meta":{"raw":{"variants":["Jump video measures height within 3.9 cm, no calibration","Gravity as ruler: video jump gives height estimate","No-cal height from jump: gravity does the work","Video jump uses earth's gravity to gauge person's height","Height from video: gravity replaces calibration"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000613,"raw_usage":{"total_tokens":2816,"prompt_tokens":875,"completion_tokens":1941,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":491,"completion_tokens_details":{"reasoning_tokens":1864}},"tokens_in":491,"tokens_out":1941,"duration_ms":14638,"temperature":1.0,"reasoning_tokens":1864,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-14T04:57:52.222480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A controlled video of a person of known height jumping at constant distance to a static camera should reproduce the paper's 3.9 cm MAE; the same person jumping while walking one meter toward the camera during the clip should produce a systematically biased estimate if the distance-consistency assumption breaks.","supporting_citations":[{"cited_title":"The Mathematical Principles of Natural Phi- losophy","cited_arxiv_id":null,"evidence_quote":"Supplies the free-fall equation of motion that yields the quadratic trajectory and the known constant $g$."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the per-frame 2D keypoint detections used to compute the center-of-mass trajectory."},{"cited_title":"Clauser, J.T","cited_arxiv_id":null,"evidence_quote":"Provides the body-segment mass ratios used to compute the center of mass from keypoints."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"RANSAC outlier rejection stabilizes the quadratic curve fit of the trajectory."},{"cited_title":"What Face and Body Shapes Can Tell About Height","cited_arxiv_id":"1805.10355","evidence_quote":"The appearance-based single-image height regression baseline whose 6.5 cm MAE the paper compares against."}],"review_version":1}