DOSS-Weight: Estimation & Data Reweighting

• DOSS-Weight is a multifaceted concept that spans non-contact 3D body weight estimation, principled data reweighting for speech deepfake detection, and observer-based weight estimation in robotics.
• It employs rigorous protocols such as 3D point cloud processing, convex hull volume computation, domain-calibrated reweighting, and sliding-mode observers to ensure accurate and real-time predictions.
• Empirical results demonstrate notable performance with metrics like a 2.8 kg MAE in medical applications, a 29% error rate reduction in deepfake detection, and convergence within 2–4 seconds in robotic mass estimation.

DOSS-Weight is a designation attributed to multiple technically unrelated methodologies across diverse research fields. Most notably, “DOSS-Weight” denotes (i) a clinically validated, non-contact 3D body mass estimation pipeline and (ii) a principled data re-weighting strategy for large-scale speech deepfake detection model training. Additionally, in mobile robotics, it serves as an implementation label for observer-based mass estimation from proprioceptive signals. This article surveys these principal usages, focusing on technical formulation, implementation procedures, hyper-parameterization, and empirical outcomes associated with each application.

1. 3D Vision-Based Body Weight Estimation in Emergency Medicine

DOSS-Weight refers to a stepwise protocol for non-invasive, real-time estimation of human body weight based on 3D surface data acquisition and geometric modeling. The system is motivated by the clinical imperative to rapidly estimate body weight for critical medication dosing, where direct measurement is impractical. Its architectural pipeline is as follows (Sonar et al., 18 Sep 2024):

1. 3D Data Acquisition: An Intel RealSense D415 or similar active stereo depth camera is rigidly mounted (1.5–2 m elevation, ~25° tilt) to capture high-resolution depth maps. Calibration exploits chessboard-based procedures and RealSense SDK routines.
2. Preprocessing: Depth thresholding, statistical outlier removal (mean neighbor distance test), and voxel-grid downsampling (5–10 mm) are used to yield artifact-free, registered 3D point clouds. RANSAC-based plane segmentation removes the supporting surface.
3. Anatomical Segmentation: Skeletal landmarks are obtained via a 3D pose estimation network (PoseNet/MediaPipe) on the RGB stream, projected into 3D, and used as seeds for limb-wise region-growing segmentation. Clusters are refined to enforce anatomical ordering and continuity.
4. Volume Calculation: The convex hull is computed for each segment using Qhull/Open3D implementations. Each hull is tetrahedralized, and its volume is calculated:

$$V(H) = \sum_{t=1}^T \frac{1}{6}|\det[v_a^t - v_0^t,\;v_b^t - v_0^t,\;v_c^t - v_0^t]|$$

Summing across all body segments, the total body volume $V_{\mathrm{total}}$ is obtained.

1. Height Estimation: Principal component analysis (PCA) is performed on the point cloud. Subject height is calculated as the projected span along the principal axis; a direct landmark-based head-to-heel distance method serves as a fallback.
2. Weight Calculation: Estimated weight is given by $W = \rho V_{\mathrm{total}}$, where $\rho$ is mean soft-tissue density (empirically ≈985 kg/m$^3$). $\rho$ is population-calibrated to minimize regression error.
3. Validation: On 50 subjects, the system attains mean absolute error (MAE) of 1.9 cm (height), 2.8 kg (weight), with 90% of estimates within ±10% of true value.

This pipeline delivers real-time (10 Hz) predictions and is robust to moderate pose/occlusion/rotation variations. Systematic sources of error include density variability, misfit of the supporting plane, depth noise, and occlusion/clothing artifacts. Recommended improvements include incorporating a statistical regression model and sensor fusion for further accuracy (Sonar et al., 18 Sep 2024).

2. Diversity-Optimized Data Reweighting for Speech Deepfake Detection

Within the speech deepfake detection (SDD) literature, “DOSS-Weight” denotes a mathematically principled framework for mixing training samples from highly imbalanced, heterogeneous domains without discard, aiming to optimize cross-domain generalization (Huang et al., 20 Dec 2025).

Mathematical Framework

Let $F$ denote the set of fake domains and $R$ the set of real domains, each with sample counts $n_f$, $n_r$ respectively. The DOSS-Weight algorithm imposes domain-level reweighting such that training draws are nearly uniform over domains, not samples, thereby mitigating domain dominance and rare attack underrepresentation.

Algorithmic steps:

1. Domain Importance Capping: For each fake domain $f$, define $u_f = \min(n_f, N_c)$ (saturation cap $N_c$). Temperature-sharpened importance: $\alpha_f = u_f^{1/\tau}$ (diversity temperature $\tau$).
2. Real Domain Aggregation: For each real domain $r$, $$u_r = \sum_{f:\text{base}(f)=r} u_f,\;\; \alpha_r = u_r^{1/\tau}$$.
3. Class Ratio Enforcement: Allocate total sampling probability mass so $P(\text{Fake}) : P(\text{Real}) = 1 : \rho$ (real/fake weighting hyperparameter $\rho$), assigning

$$w_f = \frac{\alpha_f}{A_F}\cdot \frac{1}{1+\rho}\,,\quad w_r = \frac{\alpha_r}{A_R} \cdot \frac{\rho}{1+\rho}$$

(where $A_F=\sum_{f\in F}\alpha_f$, $A_R=\sum_{r\in R}\alpha_r$, rescaled so $\sum_d w_d=1$).

1. Sample-Level Weights: For WeightedRandomSampler training, per-sample weight $w_i = w_{d_i}/n_{d_i}$, with $d_i$ being the sample’s domain.
2. Hyperparameterization: $N_c$ (empirically ≈2,500 samples) and $\tau$ ($\sim$5) are tuned via ablation for optimal error-rate (EER) reduction; $\rho$ ($0.25$) matches a realistic class proportion.

Empirical Outcomes

DOSS-Weight achieves substantial improvements:

• On a 12k-hour curation—18 real, 332 fake domains—XLS-R models trained with DOSS-Weight reduce error rates by 29% vs. naive aggregation and 13% vs. DOSS-Select (sample culling) (Huang et al., 20 Dec 2025).
• Using XLS-R-1B, DOSS-Weight achieves 1.65% EER (state-of-the-art on 6/8 benchmarks) and 96.01% accuracy on the hardest commercial API deepfake challenge, outperforming larger models trained on naive pools with 6× the data.
• Relative to naive concatenation, DOSS-Weight offers increased data and model efficiency, leading to predictable, generalization-focused sample selection.

Practical Workflow

Recommended usage:

• Curate domain inventory; select $N_c$ to match diminishing returns; increase $\tau$ to promote uniformity; calibrate $\rho$ for desired class weighting; and use sample-level weights with standard WeightedRandomSampler (Huang et al., 20 Dec 2025).

3. Sliding-Mode Observer-Based Weight Estimation in Robotic Manipulation

In mobile robotics, “DOSS-Weight” is the adopted label for the weight estimation package built around sliding-mode observers (SMO) deployed as standalone ROS nodes (Negrete et al., 2020). Here, mass (object weight) is treated as an unknown external disturbance.

• The robot arm is dynamically modeled as

$$M(q)\ddot{q} + C(q,\dot{q})\dot{q} + G(q) = \tau + f_w$$

with $f_w$ encoding the fault torque due to load.

• A super-twisting SMO is implemented:

$$\dot{\hat x}_1 = \hat{x}_2 + z_1\quad;\quad \dot{\hat x}_2=f(\cdot)+z_2$$

with injection $$z_1 = \lambda|q - \hat q|^{1/2} \text{sign}(q-\hat q)$$, $z_2 = \alpha \text{sign}(q-\hat q)$, $(\lambda=6.0,\alpha=4.2)$.

• The equivalent injection $z_{2,\text{eq}}$ approximates the unknown load. After Butterworth filtering, the relevant joint torque disturbance is used to solve for the object’s mass:

$\hat m_o = -\tau_6/(g l_4 \sin\theta)$

Estimation converges within 2–4 seconds to within 10% of the true value in both simulation and hardware, subject to modeling and sensor-limitations (Negrete et al., 2020).

4. Comparative Table of DOSS-Weight Usages

Domain	Purpose	Core Method and Metric
3D Vision/Medicine	Non-contact body weight estimation	Convex hull + density model MAE, RMSE (Sonar et al., 18 Sep 2024)
Speech Deepfake Detection	Generalizable model training via dataset reweighting	Entropy-motivated sampling, EER/ACC (Huang et al., 20 Dec 2025)
Robotics	Object mass estimation during manipulation	Sliding-mode observer, time-to-convergence (Negrete et al., 2020)

5. Cross-Domain Significance and Limitations

Although the specific mathematical and algorithmic details diverge sharply across the domains, the common thread is the use of weighting—either of geometric primitives, domain contributions, or observer signal components—to achieve robustness or accuracy when direct measurement or naive aggregation is suboptimal.

In 3D body weight estimation, limitations center around anthropometric and sensor idiosyncrasies. For DOSS-Weight in SDD, performance relies on judicious hyper-parameter selection but does not require model-based validation or data discard. In robotic mass estimation, core error sources include mechanical model uncertainty and pose-dependent mapping of disturbance to mass (Sonar et al., 18 Sep 2024, Huang et al., 20 Dec 2025, Negrete et al., 2020).

A plausible implication is that despite vastly different contexts, the DOSS-Weight label signifies a principled intervention at the level of data curation, measurement, or signal reconstruction, targeting generalization and reliability in settings where naive, unweighted procedures are inferior.