HandAnthro: Automated Hand Anthropometry from a Single Image

Fan Zhou; Shuairan Chen; Mengying Zhang; Yulin Wu; Sadegh Jafari; Sixing Yu; Rui Li; Ali Jannesari

cs.CVSep 29, 2026

HandAnthro: Automated Hand Anthropometry from a Single Image

Authors: Fan Zhou, Shuairan Chen, Mengying Zhang, Yulin Wu, Sadegh Jafari, Sixing Yu, Rui Li, Ali Jannesari, +1 more

Organizations: Iowa State University · Microsoft

Abstract

Hand anthropometry supports protective-glove design, but existing measurement methods often require trained operators, specialized hardware, or manual landmarking. We present HandAnthro, which estimates 44 projected hand dimensions from a smartphone photograph of a palm-up hand on US letter-size paper. The pipeline reconstructs wrist-occluded paper boundaries for rectification, whitens non-hand pixels, and refines 41 anthropometry-specific landmarks from a fine-tuned You Only Look Once (YOLO) pose model using image-specific geometry and contours. Controlled evaluation comprised 720 captures from 45 held-out participants, each contributing 16 images across two smartphones, two backgrounds, two angles, and two nominal illumination settings. HandAnthro produced complete outputs for 704 captures (97.8%); among these, mean absolute error (MAE) was 3.80 mm per dimension against two trained operators' caliper measurements. Regional MAEs were 2.48 mm for non-thumb fingers, 6.04 mm for thumbs, and 6.17 mm for palm and wrist. In a researcher-assisted mobile-app pilot, automated batch processing returned all 44 dimensions for 260 of 268 retained, researcher-screened firefighter images (97.0%). A descriptive, unpaired comparison with an independent national firefighter reference yielded a mean absolute difference of 2.40 mm across 28 sex-by-dimension group-mean contrasts. These results characterize controlled measurement performance and researcher-assisted field feasibility for future distributed hand-anthropometry studies.

Figures & tables

Method	Automation and equipment requirements	Study sample; dimensions	Reference	Hand-dimension MAE (mm)
Han and Park ( 2016 )	Automated ; required equipment: flatbed scanner and enclosure	11 people; 17 tabulated dimensions	Live-hand calipers	—
Kaashki et al. ( 2022 )	Automated ; required equipment: depth sensor and tablet	20 people; 11 real-hand dimensions	Anthropometrist; instrument unspecified	4.5
Nguyen, Le, and La ( 2025 )	Automated ; required equipment: industrial camera, light box, ring light, and calibration target	539 people; 24 dimensions	Gauge-block calibration	—
HandAnthro	Automated ; required equipment: smartphone and known-size paper	45 people; 44 dimensions	Mean of two operators’ caliper readings	3.80 overall; 2.48 non-thumb fingers; 6.04 thumb; 6.17 palm/wrist

Table 1: Selected hand-anthropometry methods. MAEs are not directly comparable across studies; see Appendix for definitions and reporting notes.

Figure 1: HandAnthro capture-to-measurement workflow.

Figure 2: Occlusion-aware PR workflow: (a) input image; (b,c) initial paper and hand masks; (d,e) construction of the inpainting region; (f) completed paper mask; (g) detected paper quadrilateral (green frame); and (h) perspective-rectified RGB image.

Figure 3: Geometry-constrained correction of groupwise rotation-and-scale misalignment.

Panel of Figure 4 — Figure 4: Perspective-rectification comparison.

Figure 5: Capture-condition accuracy and stability in 704 completed captures from 45 participants. (a) Caliper-referenced MAE averaged across 44 dimensions for each condition. (b) Absolute differences between the two condition-specific mean predictions, computed for each participant and dimension and then averaged across participants and dimensions. Both panels summarize completed outputs; seven participants had incomplete capture-condition grids.

Appendix figures & tables27 assets

Supplementary material from the paper’s appendix.

Appendix

Panel of Figure 7 — Figure S1: One author’s hand under the eight test-set conditions (two devices, two backgrounds, and two capture angles; illumination pooled). Fingerprints are removed.

Panel of Figure 8 — Figure S2: Controlled capture rig: the left clamp provides the top-down view and the right clamp the oblique view. Distances are in centimeters; the measured oblique configuration is approximately $25^{\circ}$ from vertical.

Figure S3: The 41 anthropometry-specific landmarks on a palmar hand, labeled by abbreviation (Table S1 ).

Region	Abbreviation range	Count	Role
Thumb	$T_{1}$ to $T_{6}$	6	Thumb-specific landmarks
Index finger	$I_{1}$ to $I_{7}$	7	Fingertip to finger root
Middle finger	$M_{1}$ to $M_{7}$	7	Fingertip to finger root
Ring finger	$R_{1}$ to $R_{7}$	7	Fingertip to finger root
Little finger	$L_{1}$ to $L_{7}$	7	Fingertip to finger root
Inter-finger webbing	$\mathrm{RO}_{1}$ to $\mathrm{RO}_{3}$	3	Root-of-finger webbing points

Appendix

Table S1: The 41 anthropometry-specific landmarks predicted directly by the YOLO model, grouped by anatomical region.

Abbrev.	Expansion	Anatomical joint (palmar crease)	Formula
$\mathrm{MFK}_{I}$	Mid First Knuckle (Index)	Index-finger DIP crease	midpoint( $I_{2}$ , $I_{3}$ )
$\mathrm{MFK}_{M}$	Mid First Knuckle (Middle)	Middle-finger DIP crease	midpoint( $M_{2}$ , $M_{3}$ )
$\mathrm{MFK}_{R}$	Mid First Knuckle (Ring)	Ring-finger DIP crease	midpoint( $R_{2}$ , $R_{3}$ )
$\mathrm{MFK}_{L}$	Mid First Knuckle (Little)	Little-finger DIP crease	midpoint( $L_{2}$ , $L_{3}$ )
$\mathrm{MFK}_{T}$	Mid First Knuckle (Thumb)	Thumb IP crease	midpoint( $T_{2}$ , $T_{3}$ )
$\mathrm{MSK}_{I}$	Mid Second Knuckle (Index)	Index-finger PIP crease	midpoint( $I_{4}$ , $I_{5}$ )

Appendix

Table S2: The 14 midpoint landmarks derived from the 41 directly predicted anthropometry-specific landmarks.

Panel of Figure 12 — Figure S4: Illustrative SAM-HQ prompts for (a) paper-mask and (b) hand-mask generation. Red points are positive prompts; blue points are negative prompts.

Figure S5: Background whitening using the warped hand mask: (a) raw hand mask; (b) detected paper quadrilateral (green); (c) rectified hand mask; (d) background-whitened image.

Item	Value
Epochs / batch size	8000 / 40
Early stopping patience	10000
Compute	CUDA GPU; dataloader workers: 8
Optimizer	auto (Ultralytics default selection)
Learning rate schedule	$\mathrm{lr}_{0}=0.001$ , $\mathrm{lrf}=0.01$
Momentum / weight decay	0.937 / $5\times 10^{-4}$

Appendix

Table S3: Key hyperparameters and settings for YOLO fine-tuning.

Figure S6: Finger-axis estimation and fingertip correction for finger $f$ .

Variant	Mean (mm)	Median (mm)	$\Delta$ MAE (mm / %)	Runtime (ms)
Ours	3.80	3.45	—	17
REMBG	4.26	3.79	$+0.46$ / $+12.0\%$	93
Background Remover	4.26	3.82	$+0.46$ / $+12.1\%$	188
CarveKit	4.71	4.48	$+0.91$ / $+23.9\%$	379

Appendix

Table S4: Background-whitening alternatives under identical PR, landmark-prediction, and PP stages. MAE is measured against calipers on the 704 PR-complete captures, all of which yielded complete outputs for every variant; runtime is BW-only mean latency over the same 704 images. Positive

\Delta

MAE indicates degradation relative to mask reuse.

Figure S7: BW-stage runtime versus caliper-referenced dimension MAE on the 704 complete captures. Runtime is measured on an NVIDIA RTX 4000 Ada Generation GPU; the horizontal axis is logarithmic. Values match Table S4 .

Figure S8: Our BW versus the three off-the-shelf baselines on representative images. REMBG, BackgroundRemover, and CarveKit often retain inter-finger shadows that blur the hand boundary, whereas our BW gives a clean separation.

Figure S9: Per-dimension MAE for $E_{16}$ across 44 dimensions (704 images, 45 participants). (a) Distribution (median 3.45 mm, mean 3.80 mm). (b) Spatial map in which each line connects a dimension’s two endpoints and is colored by its MAE magnitude.

ID	Measurement	Endpoints	Mean $\pm$ SD	Median	$P_{95}$	Range
Thumb (5)
D1	Thumb IP breadth	$T_{2}\!\to T_{3}$	$3.84\pm 1.46$	3.77	6.57	0.98–10.01
D2	Thumb MCP breadth	$T_{4}\!\to T_{5}$	$6.51\pm 3.71$	6.45	12.63	0.06–18.17
D3	Thumb distal-segment length	$T_{1}\!\to\mathrm{MFK}_{T}$	$5.69\pm 3.09$	5.36	10.87	0.00–14.03
D4	Thumb proximal-segment length	$\mathrm{MFK}_{T}\!\to T_{6}$	$7.82\pm 4.43$	7.57	15.65	0.05–23.46
D5	Thumb total length	$T_{1}\!\to T_{6}$	$6.32\pm 4.08$	6.01	13.49	0.04–20.18

Appendix

Table S5: Definitions and per-dimension absolute-error statistics for

E_{16}

on the 704 complete outputs. Endpoints use the notation in Tables S1 – S2 ; all statistics are in millimeters, and the Range column reports minimum and maximum errors.

Metric	$E_{15}$ (A2)	$E_{16}$ (A3)	$\%\Delta$
KP Mean (px)	11.18	7.83	$-29.94\%$
KP Std (px)	8.73	6.74	$-22.77\%$
Median error (px)	8.87	6.12	$-30.98\%$
IQR (px)	9.13	6.26	$-31.47\%$
$P(\text{error}\leq 5\,\text{px})$	21.52%	39.02%	$+81.3\%$
$P(\text{error}\leq 10\,\text{px})$	57.56%	75.77%	$+31.6\%$

Appendix

Table S6: Pixel-level comparison of

E_{15}

(A2, no PP) and

E_{16}

(A3, with PP) on 45 annotated test images (1,845 keypoint pairs).

\%\Delta=(E_{16}-E_{15})/E_{15}

Group	$E_{15}$ (A2)	$E_{16}$ (A3)	$\%\Delta$
Index	8.57	5.59	$-34.8\%$
Middle	8.70	6.05	$-30.5\%$
Ring	10.81	6.90	$-36.1\%$
Little	13.13	7.35	$-44.0\%$
Thumb	17.35	14.54	$-16.2\%$
Root	5.91	5.85	$-1.0\%$

Appendix

Table S7: Mean keypoint pixel error (px) by anatomical group for

E_{15}

(A2) and

E_{16}

(A3) on 45 real test images.

Panel of Figure 23 — Figure S10: Representative outputs across the progressive ablation configurations.

Group ( $n$ )	A2	A3	$\Delta$ [95% CI]
Overall (44)	3.707	3.804	$+0.097\;[0.002,0.193]$
Finger length (19)	3.196	3.097	$-0.100\;[-0.235,0.038]$
Finger width (14)	2.722	2.910	$+0.188\;[0.035,0.340]$
Palm/wrist (11)	5.842	6.165	$+0.323\;[0.167,0.494]$

Appendix

Table S8: Paired PP ablation on 704 common completed captures from 45 participants. A2 and A3 columns report MAE;

n

counts dimensions. All values are in millimeters, and negative

\Delta

indicates improvement. Confidence intervals use participant-cluster resampling.

	Mean absolute error				Signed bias		$P_{95}$ absolute error
ID	A2	A3	$\Delta$	95% CI for $\Delta$	A2	A3	A2	A3
D1	2.709	3.843	$+1.134$	$[0.775,1.504]$	$+2.700$	$+3.809$	4.830	6.565
D2	8.450	6.509	$-1.941$	$[-2.389,-1.501]$	$+8.426$	$+6.324$	15.552	12.629
D3	3.280	5.691	$+2.410$	$[1.967,2.878]$	$-2.600$	$-5.540$	7.641	10.866
D4	11.083	7.819	$-3.264$	$[-4.055,-2.474]$	$+10.884$	$+7.318$	18.193	15.655
D5	9.927	6.317	$-3.610$	$[-4.720,-2.540]$	$+9.794$	$+3.336$	17.745	13.486

Appendix

Table S9: Complete dimension-level PP comparison on the same 704 captures. IDs follow Table S5 , and all values are in millimeters.

\Delta=\mathrm{MAE}_{\mathrm{A3}}-\mathrm{MAE}_{\mathrm{A2}}

; negative values indicate improvement. The paired participant-bootstrap intervals are exploratory, without multiplicity adjustment. Signed bias is the mean signed error (prediction minus caliper reference);

P_{95}

is the per-dimension 95th percentile of per-capture absolute errors, using linear interpolation.

ID	YOLO	Train aug.	Case	Pose (Train)	Pose (Val)
Without training data augmentation
E1	v8	no	A0	7.29355	3.42462
E2	v8	no	A1	0.94881	0.78883
E3	v8	no	A2	0.93568	0.83675
E4	v8	no	A3	–	–
E5	v11	no	A0	9.02002	8.49234

Appendix

Table S10: Training and validation pose loss for all 16 configurations.

Factor	$A$ / $B$	$N_{A}$ / $N_{B}$	MAE A / MAE B	$\Delta$
Device	Android / iPhone	350 / 354	3.83 / 3.78	$+0.05$
Background	Complex / Simple	354 / 350	3.80 / 3.80	$-0.00$
Angle	Top-down / Oblique	346 / 358	4.18 / 3.44	$+0.74$

Appendix

Table S11: Descriptive accuracy by capture factor among complete outputs.

N_{A}

and

N_{B}

count images, not independent participants;

\Delta

MAE

=

MAE

{}_{A}-

MAE B , and all MAE values are in millimeters.

		Within-person CV (%)		Mean absolute condition-mean difference (mm)
ID	Measurement	Median	$P_{95}$	Angle	Smartphone	Background
D1	Thumb IP breadth	3.16	7.15	1.02	0.37	0.62
D2	Thumb MCP breadth	4.39	8.42	1.88	0.84	1.23
D3	Thumb distal-segment length	9.43	12.36	3.87	0.76	1.42
D4	Thumb proximal-segment length	7.14	9.80	3.77	0.85	1.52
D5	Thumb total length	7.93	10.17	7.73	1.28	2.24

Appendix

Table S12: Within-person capture variability across 44 dimensions. For each dimension, median and

P_{95}

summarize the CVs of 45 participants; each condition-mean difference is the average across participants of the absolute difference between their two condition-specific mean predictions. Summaries use the 704 completed captures. Dimension IDs and endpoints follow Table S5 .

Pattern	Dimensions	$\Delta_{d}$ (mm)
Top-down worse	Wrist-referenced dimensions (D38–D44)	$+2.92$ to $+4.28$
	Middle/ring total lengths (D19, D26)	$+3.19$ , $+2.97$
	Thumb proximal/total lengths (D4, D5)	$+2.89$ , $+2.22$
Oblique worse	Thumb distal length (D3)	$-3.74$
	Index/middle/ring DIP and PIP breadths (D6, D7, D13, D14, D20, D21)	$-1.09$ to $-0.58$
Near zero	12 of 44 dimensions	$\|\Delta_{d}\|\leq 0.50^{\dagger}$

Appendix

Table S13: Largest and recurrent dimension-specific capture-angle contrasts. Positive

\Delta_{d}

indicates higher top-down MAE; all contrasts are descriptive. Dimension definitions appear in Table S5 .

Stage	CPU (M3)	GPU (RTX 4000 Ada)
Total execution time	$26.36\pm 5.59$	$3.86\pm 0.61$
Perspective rectification	$24.98\pm 5.25$	$2.30\pm 0.38$
Background whitening	$\phantom{0}0.02\pm 0.01$	$0.02\pm 0.01$
Inference and PP	$\phantom{0}1.16\pm 0.37$	$1.51\pm 0.31$

Appendix

Table S14: Per-image stage runtime (mean

\pm

SD, seconds). CPU: Apple M3, 16 successful images from a 20-image batch; GPU: RTX 4000 Ada, 704 complete outputs. BW reports mask-to-white composition only; cached-mask I/O is excluded.

Figure S11: Bland–Altman plot of signed relative differences between the two trained operators ( $N=1{,}980$ across 45 hands and 44 dimensions). Horizontal lines indicate the pooled mean bias and 95% limits of agreement.

Reference dimension	HandAnthro	Note
Hand length	D40 + D19	palm length + middle-finger length
Hand breadth	D34	across the metacarpals
Palm length	D40	wrist-midpoint to middle MCP
Palm breadth	D34	proxy (no palm-crease breadth)
Thumb length	D5	tip to thumb root
Thumb breadth	D1	IP-joint breadth

Appendix

Table S15: Mapping from Hsiao (2015) Table 2 dimensions to HandAnthro outputs.

Dimension	Ours	Ref	$\Delta$	$\Delta\%$	$g$	FDR
Hand length	204.9 (10.6)	197.6 (9.3)	$+7.33$	$+3.7$	$+0.77$	yes
Hand breadth	100.8 (6.0)	97.2 (4.6)	$+3.58$	$+3.7$	$+0.73$	yes
Palm length	121.5 (6.0)	113.8 (5.8)	$+7.75$	$+6.8$	$+1.33$	yes
Palm breadth	100.8 (6.0)	96.0 (4.6)	$+4.78$	$+5.0$	$+0.98$	yes
Thumb length	76.0 (5.8)	70.8 (4.3)	$+5.18$	$+7.3$	$+1.12$	yes
Thumb breadth	27.7 (7.4)	24.4 (1.6)	$+3.31$	$+13.6$	$+0.94$	yes

Appendix

Table S16: Sex-specific comparison of 260 successful dominant-hand palmar cases (197 male and 63 female) with Hsiao et al. Ours and Ref are mean (SD) in millimeters;

\Delta=

Ours

-

Ref,

g

is Hedges’

g

, and FDR indicates Benjamini–Hochberg significance at

q<0.05

Explore similar work

ScaleHP: Estimating Hand Pose in Metric Space
Jun 24, 2026Ruitao Jing, Xingyu Chen, Hongyang Li +33D Hand Hand
Tsinghua University · Visincept · International Digital Economy Academy (IDEA Research) +2
VEPHand: View-Efficient Photometric Hand Performance Capture at Scale
Jun 14, 2026Zhengyang Shen, Kai-Hung Chang, Erroll Wood +183D Hand Motion Capture
Google XR
REACH: Hand Pose Estimation from Room Corners
May 21, 2026Shu Nakamura, Ryo Kawahara, Genki Kinoshita +43D Hand
Graduate School of Informatics, Kyoto University · RIKEN · Kyoto Institute of Technology

Jun 24, 2026, cs.CVJ/K move · Enter open · S save

ScaleHP: Estimating Hand Pose in Metric Space

Ruitao Jing, Xingyu Chen, Hongyang Li, Qing Jiang+2

Tsinghua University · Visincept · International Digital Economy Academy (IDEA Research)+2

3D Hand Hand Stable Anthropometric Priors Metric Spaces

Open paper AskPDF arXiv

Accurate metric-space hand pose estimation (HPE) is essential for immersive human-computer interaction and robotics. However, most existing methods predict poses in a root-relative coordinate system and cannot estimate the hand in absolute metric scale. In this work, we observe that the intrinsic proportional relationships among human hand bones encode stable anthropometric priors that implicitly correlate with the overall metric size of the hand. Leveraging this insight, we present ScaleHP, an end-to-end one-stage hand pose estimation framework that bypasses fragile extrinsic depth modules to recover the hand in metric space. ScaleHP employs a transformer-based decoder with a novel scale token to fuse multi-scale morphological and appearance features. By solving for metric coordinates through a perspective-constrained least-squares approach, we achieve high-precision pose estimation in the camera coordinate system. ScaleHP delivers state-of-the-art performance, including 35.8 CS-MPJPE on FreiHand and 4.6/5.9 PA-MPJPE on DexYCB and HO3Dv3. These results demonstrate that internal biological constraints significantly reduce relative geometry and absolute metric errors, offering a robust solution for generalized, real-world hand tracking.

Abstract

Explore similar work

ScaleHP: Estimating Hand Pose in Metric Space

VEPHand: View-Efficient Photometric Hand Performance Capture at Scale

REACH: Hand Pose Estimation from Room Corners