Body Segment Volume Estimation

• BSV is the process of estimating whole-body, segmented, or tissue-specific volumes from imaging data such as RGB, CT, MRI, and depth scans.
• Techniques range from direct geometric reconstruction and segmentation-based volumetry to deep learning methods for inferring volumes from external shapes.
• Applications span emergency medicine, cardiometabolic risk, and biomechanics, while challenges include sensor calibration, validation consistency, and methodological ambiguities.

Searching arXiv for recent and canonical papers on body segment volume estimation and closely related volumetry methods. Body Segment Volume Estimation (BSV) denotes the estimation of whole-body volume, individual body-part volumes, or anatomically defined tissue-compartment volumes from visual, depth, surface-mesh, CT, or MRI data. In current arXiv literature, the term spans monocular RGB regression of 14 body-part volumes and total body volume, dual RGB-D reconstruction of whole-body and segment volumes, depth-derived convex-hull aggregation, CT and MRI mask-based volumetry, and graph-based prediction of internal adipose-tissue compartments from body surface meshes (Leinen et al., 2021, Bassani et al., 12 Sep 2025, Sonar et al., 2024, Ma et al., 2021, Liao et al., 2017, Mueller et al., 2023). Across these formulations, BSV is not a single algorithmic problem but a family of estimation tasks whose outputs may be geometric body-part volumes, tissue volumes in $cm^3$, or uncertainty-calibrated volumetric intervals.

1. Scope and task variants

The literature uses BSV in both a strict and an extended sense. In the strict anthropometric sense, the target is the volume of the whole body or explicit anatomical parts such as head, torso, arms, forearms, hands, thighs, shins, and feet. In the extended body-composition sense, the targets are tissue compartments such as skeletal muscle, bone, subcutaneous adipose tissue (SAT), visceral adipose tissue (VAT), abdominal subcutaneous adipose tissue (ASAT), or lumbar vertebral-body volumes. A further extension is crowd-level aggregation, where the objective is the collective body volume of all people visible in an RGB image (Collorone et al., 3 Jan 2025).

Paradigm	Input	Output
Monocular RGB regression	One RGB image plus body height	Volumes of 14 body parts and total body volume
RGB-D / depth reconstruction	Two RGB-D cameras or one overhead depth camera	Whole-body and segment volumes
Surface-mesh inference	Triangulated body surface mesh	VAT and ASAT volumes
CT / MRI volumetry	3D CT or MRI slices	Tissue volumes, vertebral volumes, LV volume
Crowd aggregation	Single RGB image of a crowd	Collective body volume

A representative formulation is VolNet, which takes one RGB image and the person’s body height and predicts the volumes of 14 body parts and the total body volume (Leinen et al., 2021). A different formulation estimates body volume and height from 3D data acquired with an overhead Intel RealSense D415 camera and uses those quantities to support body-weight estimation in emergency medicine (Sonar et al., 2024). In CT-based work, whole-body or regional segmentation directly yields volumes for skeletal muscle, bone, SAT, and VAT, often in vertebra-indexed slabs or in chest–abdomen–pelvis subvolumes (Ma et al., 2021, Chen et al., 13 Feb 2025).

A common misconception is that BSV always means direct geometric reconstruction of external body shape. The literature shows three distinct targets: direct geometric volume, segmentation-derived tissue volume, and inference of internal volume from external shape proxies. Mesh-based VAT/ASAT prediction is a whole-graph regression problem, not a direct measurement of visible segment geometry (Mueller et al., 2023). Conversely, CT workflows are direct volumetry from labeled voxels rather than regression from surface appearance (Koitka et al., 2020).

2. Geometric pipelines and explicit volume computation

A large class of BSV methods is geometric and reconstruction-driven. In the depth-camera pipeline for emergency medicine, the RealSense D415 is positioned directly above the subject and captures depth maps at $848 \times 480$ pixels and 30 FPS. Background is removed with the DepthQualityTool SDK, body-part segmentation is performed with models such as MobileNet and ResNet, a polygon file is constructed from the depth-derived data, and filtering and smoothing are applied before volume is computed from convex hulls on multiple body sections (Sonar et al., 2024). Rather than computing one convex hull over the entire body, the method subdivides the point cloud into anatomical regions and sums their volumes: $V_{\text{total}}=\sum_{i=1}^{n} V_i$ where $V_i$ is the convex-hull-derived volume of the $i$-th body segment. Height is derived from the bounding-box extent,

$H = Y_{\max} - Y_{\min}.$

A related but more explicitly anthropometric system uses two Intel RealSense L515 LiDAR RGB-D cameras placed front and back, performs extrinsic calibration with an ArUco marker, detects landmarks with MediaPipe Holistic, segments body parts with BodyPix, cleans and merges point clouds, and reconstructs a watertight mesh through enhanced As-Rigid-As-Possible (ARAP) non-rigid registration (Bassani et al., 12 Sep 2025). The ARAP energy is

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$

with cotangent weights

$$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$

The paper adds a regularization term scaled by total mesh area $A$ to prevent collapse in lateral gap regions, then computes whole-body and isolated segment volumes from the closed meshes. On FAUST, the reported whole-body mean Relative Volume Error (RVE) is approximately $1.23\%$ in NoEr and $848 \times 480$0 in L515; box experiments report RVE below $848 \times 480$1 (Bassani et al., 12 Sep 2025).

Low-cost 3D scanning has also been used to estimate segment masses for biomechanics. A whole-body Sense™ 3D scanner produces an STL mesh, markers define anatomical cuts, the mesh is dismembered into 15 body parts, and segment masses are estimated via

$848 \times 480$2

The study reports that the mass error in one segment can be as large as $848 \times 480$3 of overall body weight and that mean influence on calculated joint moment varies from $848 \times 480$4 to $848 \times 480$5 in 18 joints (Chang et al., 2018). This indicates that segment volume estimation is not merely descriptive geometry; it directly changes downstream inverse-dynamics outputs.

3. Learning-based inference from RGB images and surface meshes

Monocular RGB BSV is exemplified by VolNet, a multi-stage architecture that predicts 2D pose, body-part segmentation, 3D pose, and finally body-part volumes from the RGB image plus body height (Leinen et al., 2021). The first three tasks use stacked hourglass subnetworks, while the volume regressor uses a ResNeXt-50 backbone with cardinality 32 and two fully connected layers. Intermediate tasks are trained with MSE for 2D and 3D pose and categorical cross-entropy for segmentation; final volume regression minimizes MSE between predicted and ground-truth body-part volumes. The synthetic SURREALvols dataset contains 372,142 images at $848 \times 480$6 resolution, with annotations for 2D pose, 3D pose, body-part segmentation, body height, and volumes of 14 body parts. On the test set, VolNet reports total-volume MAPE of $848 \times 480$7 and a success rate within $848 \times 480$8 tolerance of approximately $848 \times 480$9, compared with BodyNet’s mean APE of $V_{\text{total}}=\sum_{i=1}^{n} V_i$0 and about $V_{\text{total}}=\sum_{i=1}^{n} V_i$1 success at $V_{\text{total}}=\sum_{i=1}^{n} V_i$2 tolerance (Leinen et al., 2021). The paper also states that removing body height causes a large drop in MAPE, making height the critical scale anchor.

Surface-mesh BSV replaces RGB appearance with explicit external geometry. A triangulated mesh $V_{\text{total}}=\sum_{i=1}^{n} V_i$3 is treated as a graph $V_{\text{total}}=\sum_{i=1}^{n} V_i$4, and a Graph Neural Network with GraphSAGE convolutions performs whole-graph regression of VAT and ASAT from the body surface (Mueller et al., 2023). The node update is

$V_{\text{total}}=\sum_{i=1}^{n} V_i$5

The architecture uses three GraphSAGE layers, batch normalization, max aggregation, and a three-layer MLP, trained for 150 epochs with Adam and shrinkage loss. Meshes are generated with marching cubes and decimated to 10,000, 5,000, 1,000, 500, 200, and 100 faces. Reported test $V_{\text{total}}=\sum_{i=1}^{n} V_i$6 scores reach up to $V_{\text{total}}=\sum_{i=1}^{n} V_i$7 for VAT and $V_{\text{total}}=\sum_{i=1}^{n} V_i$8 for ASAT, with GNN training times of about 8–38 min depending on decimation; a 1,000-face mesh is highlighted as a particularly good trade-off with VAT $V_{\text{total}}=\sum_{i=1}^{n} V_i$9, ASAT $V_i$0, and training time $V_i$1 minutes (Mueller et al., 2023).

Crowd Volume Estimation (CVE) generalizes BSV from one person to all visible people in a scene. The task is defined as estimating the undergarment total body volume occupied by human bodies in a scene from a single RGB image, with objective

$V_i$2

The ANTHROPOS-V benchmark provides 768 Full HD videos at 30 fps across 384 diverse scenarios, with per-person total body volume, SMPL shape parameters, keypoints, and per-part volume annotations (Collorone et al., 3 Jan 2025). STEERER-V replaces head-centered density supervision with Per-Part Volume Density Maps, distributing torso volume across five torso keypoints and using a more spatially distributed supervision signal. This suggests that segment-level supervision remains useful even when the final task is crowd-level aggregation.

4. Segmentation-based medical volumetry

Medical BSV is dominated by segmentation-derived volumetry. In whole-body CT body-composition analysis, automated 3D segmentation labels skeletal muscle, bone, SAT, and VAT across the scanned body, and volumes are then computed from segmented voxels in $V_i$3 (Ma et al., 2021). Validation on 50 evaluation CT volumes reports average Dice similarity coefficients of 0.980 for bone, 0.974 for skeletal muscle, 0.986 for SAT, and 0.960 for VAT, with Dice defined as

$V_i$4

The same study also reports vertebra-indexed regional slab volumes from 21 vertebral levels, C4–C7, T1–T12, and L1–L5, and frames this expansion from single-slice proxies to whole-body 3D tissue volumetric distribution as “extended body composition” (Ma et al., 2021).

A publicly accessible CT pipeline segments skeletal muscle, SAT, VAT, and muscular fat across the chest, abdomen, and pelvis using a 2D nnU-Net with ResEnc presets / nnU-Net ResEnc XL-style configuration. TotalSegmentator localizes T12, L3, and L4, enabling 2D measurement at the L3 slice and 3D measurement over the T12-to-L4 subvolume (Chen et al., 13 Feb 2025). The overlap metric is

$V_i$5

and the error for derived body-composition measurements is

$V_i$6

On the internal dataset, average Dice is $V_i$7 at L3, $V_i$8 for T12–L4, and $V_i$9 across all slices, with average MRAE of $i$0, $i$1, and $i$2; on the external SAROS dataset, average Dice is $i$3 at L3, $i$4 for T12–L4, and $i$5 across all slices, with average MRAE of $i$6, $i$7, and $i$8 (Chen et al., 13 Feb 2025). The paper explicitly notes that its “volume” estimates are derived from stacked axial slices and voxel geometry rather than a full dedicated 3D body-segment volumetric model.

Whole-abdomen CT analysis likewise combines 3D semantic segmentation with HU-based tissue subclassification. A multi-resolution U-Net 3D predicts abdominal cavity, bones, muscle, subcutaneous tissue, and thoracic cavity, after which muscle is identified by HU between $i$9 and $H = Y_{\max} - Y_{\min}.$0 and adipose tissue by HU between $H = Y_{\max} - Y_{\min}.$1 and $H = Y_{\max} - Y_{\min}.$2; adipose tissue inside abdominal cavity is VAT, and adipose tissue inside subcutaneous tissue is SAT (Koitka et al., 2020). The combined loss is

$H = Y_{\max} - Y_{\min}.$3

The best model reports test-set Dice scores of 0.9735 for abdominal cavity, 0.9423 for bones, 0.9334 for muscle, 0.9623 for subcutaneous tissue, and 0.9652 for thoracic cavity, with average 0.9553; ICC values are 0.999 for SAT, 0.998 for VAT, and 0.991 for muscle (Koitka et al., 2020).

MRI-based BSV follows the same logic of slice-wise segmentation plus geometric integration. For LV volumetry, an ROI detector localizes the chamber, a hypercolumns fully convolutional network segments it, and physical area is integrated along the short-axis stack. The interval volume is modeled as a truncated circular cone,

$H = Y_{\max} - Y_{\min}.$4

with total volume

$H = Y_{\max} - Y_{\min}.$5

An end-to-end differentiable approximation replaces the square-root term during training, and the final system ranked 4th on the Kaggle challenge with mean CRPS 0.010666 (Liao et al., 2017). Vertebral-body volumetry from spiral CT is another specialized instance: a coarse-to-fine pipeline combining deformable balloon segmentation, multiseeded volume growing, automatic pedicle cut, and trabecular compartment extraction reports accuracy errors below $H = Y_{\max} - Y_{\min}.$6 for volume in phantom data and intra- and inter-operator precision below $H = Y_{\max} - Y_{\min}.$7 in patient data (Mastmeyer et al., 2017).

5. Statistical estimation, uncertainty, and evaluation

The field includes both deterministic volumetry and uncertainty-aware estimation. A statistical baseline for convex bodies studies points uniformly distributed over a convex set $H = Y_{\max} - Y_{\min}.$8 under a Poisson point process model and shows that the convex hull

$H = Y_{\max} - Y_{\min}.$9

is a sufficient and complete statistic for the volume problem when $$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$0 is known (Baldin et al., 2015). The oracle estimator is

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$1

it is exactly unbiased, UMVU, and has variance

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$2

For unknown $$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$3, the practical estimator

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$4

is asymptotically unbiased with exponentially small bias (Baldin et al., 2015). Although this work is not an anthropometric imaging pipeline, it formalizes a core geometric issue in hull-based volume estimation: the raw hull underestimates true volume because of missing boundary mass.

Uncertainty quantification in medical volumetry has been addressed with conformal prediction. Given a predicted lower and upper volume $$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$5 and true volume $$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$6, the conformal score is

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$7

the calibration quantile is

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$8

and the predictive interval is

$$E_{ARAP} = \sum_{i=1}^{n}w_{i}\sum_{j\in\mathcal{N}(i)} w_{ij}\parallel(p_{i}'-p_{j}')-R_{i}(p_{i}-p_{j})\parallel^2$$9

Under exchangeability, marginal coverage satisfies

$$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$0

Weighted conformal prediction then reweights calibration samples through a density ratio estimated from compressed latent representations of a MONAI Dynamic U-Net (Lambert et al., 2024). On shifted synthetic volumetry, standard conformal prediction drops to $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$1 coverage while W-Oracle reaches $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$2 and W-Latent $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$3; on shifted brain-tumor volumetry with target coverage $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$4, standard conformal prediction miscovers at roughly $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$5 for all three tumor subregions, while weighted methods improve coverage (Lambert et al., 2024).

Evaluation protocols vary strongly across BSV subfields. RGB body-volume regression reports MAPE, AE, and APE (Leinen et al., 2021); RGB-D reconstruction reports RVE and, in real acquisitions, Relative Mass Error (Bassani et al., 12 Sep 2025); surface-mesh adipose prediction reports $$w_{ij} = \frac{1}{2}(cot\alpha_{ij}+cot\beta_{ij}).$$6 and training time (Mueller et al., 2023); CT segmentation uses Dice, ICC, and MRAE (Ma et al., 2021, Chen et al., 13 Feb 2025, Koitka et al., 2020); crowd-volume estimation uses MAE and PP-MAE (Collorone et al., 3 Jan 2025). This suggests that cross-paper comparisons are modality-specific unless the measurement target, anatomical scope, and evaluation metric are aligned.

6. Applications, limitations, and bibliographic ambiguities

The application range of BSV is broad. In emergency medicine, body volume and height are intended to support body-weight estimation for proper dosing of weight-based medications when direct measurement is impractical (Sonar et al., 2024). In cardiometabolic risk assessment, VAT and ASAT are treated as stronger indicators than BMI or waist circumference (Mueller et al., 2023). In oncology and precision medicine, whole-body and vertebra-indexed tissue volumes are proposed as patient-specific features for chemotherapy toxicity prediction, surgical outcome prediction, survival modeling, nutritional assessment, aging and cachexia phenotyping, and longitudinal monitoring (Ma et al., 2021). In biomechanics, subject-specific segment volumes and densities alter inverse-dynamics outputs in OpenSim (Chang et al., 2018). In osteoporosis, lumbar vertebral-body volume and trabecular VOIs support combined geometry and BMD assessment (Mastmeyer et al., 2017). In ergonomics, sports biomechanics, and clinical monitoring, dual RGB-D BSV enables proximal/distal volume ratios from a portable setup (Bassani et al., 12 Sep 2025). Crowd-scale estimation further connects body volume to infrastructure stress assessment and weight-balance analysis (Collorone et al., 3 Jan 2025).

The main limitations are equally heterogeneous. Monocular RGB methods rely on known body height, synthetic training data, and viewpoints represented in the training distribution; performance degrades for views from above, from below, and some back views (Leinen et al., 2021). RGB-D reconstruction requires the subject to stand still, depends on good calibration and segmentation, and shows greater variability in upper-arm estimates; head, hands, and feet were discarded from the final two-camera volume analysis (Bassani et al., 12 Sep 2025). CT methods face limited validation set sizes, inter-slice inconsistency, partial field-of-view scans, artifact sensitivity, and the fact that some “volume” estimates are assembled from independently segmented axial slices rather than full 3D geometric reconstruction (Ma et al., 2021, Chen et al., 13 Feb 2025, Koitka et al., 2020). Mesh-based inference from external surface shape remains a proxy model for internal tissue distribution (Mueller et al., 2023). Conformal intervals improve coverage under shift, but interval width increases and the method requires unlabeled test samples for weight estimation (Lambert et al., 2024).

A further literature issue is bibliographic ambiguity. The record "Lightweight Neural Framework for Robust 3D Volume and Surface Estimation from Multi-View Images" (Farchione et al., 22 Jun 2026), despite its title and abstract, is described in the provided details as a CVPR-style author-response / rebuttal LaTeX template containing no algorithm, no network, no dataset, no experiments, and no equations for volume estimation. As a source for BSV, it has essentially no direct relevance. This is a useful corrective: in BSV, paper titles and bibliographic context are not sufficient evidence of methodological content.

Overall, BSV is best understood as a set of volumetric estimation paradigms unified by a common output type rather than by a common architecture. Some methods estimate explicit segment geometry from reconstructed meshes, some infer absolute volume from monocular appearance using scale anchors, some compute tissue volumes directly from voxel masks, and some attach calibrated intervals to those measurements. The unifying research problem is quantitative volumetry of anatomically meaningful body structures under practical constraints of sensing modality, annotation regime, and deployment environment.