Advanced Acceptance Score: A Holistic Measure for Biometric Quantification

Aman Verma; Seshan Srirangarajan; Sumantra Dutta Roy

cs.CVFeb 17, 2026

Advanced Acceptance Score: A Holistic Measure for Biometric Quantification

Authors: Aman Verma, Seshan Srirangarajan, Sumantra Dutta Roy

Organizations: Dept. of Electrical Engineering, IIT Delhi, India · Dept. of Electrical Engineering and Bharti School of Telecommunications at IIT Delhi, India

Abstract

Quantifying biometric characteristics within hand gestures involve derivation of fitness scores from a gesture and identity aware feature space. However, evaluating the quality of these scores remains an open question. Existing biometric capacity estimation literature relies upon error rates. But these rates do not indicate goodness of scores. Thus, in this manuscript we present an exhaustive set of evaluation measures. We firstly identify ranking order and relevance of output scores as the primary basis for evaluation. In particular, we consider both rank deviation as well as rewards for: (i) higher scores of high ranked gestures and (ii) lower scores of low ranked gestures. We also compensate for correspondence between trends of output and ground truth scores. Finally, we account for disentanglement between identity features of gestures as a discounting factor. Integrating these elements with adequate weighting, we formulate advanced acceptance score as a holistic evaluation measure. To assess effectivity of the proposed we perform in-depth experimentation over three datasets with five state-of-the-art (SOTA) models. Results show that the optimal score selected with our measure is more appropriate than existing other measures. Also, our proposed measure depicts correlation with existing measures. This further validates its reliability. We have made our code public.

Figures & tables

Fig. 1: Comparison of the proposed $nA_{r}^{*}(\Delta)$ with existing ranking and retrieval measures. We compare the optimal score suggested by the measures in terms of four design criteria: (i) deviation in ranking order of gestures, (ii) entanglement between biometric traits of different gestures, (iii) trend deviation, and (iv) quality or relevance of quantification estimates. For optimal score, the first three are expected to be minimal, while relevance is to be maximized. We have normalized the values in the range between $[0,1]$ . This analysis reveals that score values selected by $nA_{r}^{*}(\Delta)$ jointly satisfies all the design criteria. This is not observed for other measures. This merit can be accredited to multi-facet and task specific evaluation.

Fig. 2: Computation of trend match distance (

\Psi

): With

\Psi

we find out if DGBQA scores of two consecutively ranked gestures have the same biometric separation as the ground truth. Firstly, all the gestures are arranged as per the ranking order in ground truth. The last gesture (here

G_{3}

) carries the highest biometric characteristics. To compute

\Psi

we traverse in two directions: (i) forward pass: along increasing order (refer (a)) and (ii) backward pass: along decreasing order (refer (b)). We utilize the difference between consecutive DGBQA scores to estimate the ground truth. The summation of difference between estimated and ground truth score from both the passes gives

\Psi

for that gesture. This is summed across all the gestures to derive

\Psi

Evaluation Measure	Soli					HandLogin					TinyRadar
Evaluation Measure	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$
$\hat{r}$	ViViT (1.5,0.5)	0.45	8.54	1.70	0.351	MViT (0.5,1.0)	0.00	4.11	2.09	0.327	ViViT (0.5,1.5)	1.27	8.39	2.24	0.488
$\mathcal{R}$	MF (1.5,0.5)	0.45	8.71	1.56	0.361	MViT (1.0,1.5)	0.00	4.12	1.25	0.284	ViViT (1.0,2.5)	1.45	8.50	2.13	0.540
$\Psi$	ViViT (1.0,0.5)	1.54	8.32	1.36	0.309	MViT (1.0,1.0)	0.50	3.95	0.82	0.373	TAM (1.0,1.0)	2.00	8.00	1.99	0.651
$C_{d}$	TPN (1.0,0.5)	1.72	8.12	2.39	0.088	TPN (1.5,2.5)	1.25	1.42	3.95	0.001	ViViT (0.5,2.5)	1.27	8.34	2.59	0.469
$A_{r}(\Delta)$ [ 6 ]	ViViT (0.5,1.0)	0.45	8.54	1.70	0.351	MViT (0.5,1.0)	0.00	4.11	2.09	0.327	MF (1.0,1.5)	1.45	8.28	2.27	0.484

TABLE I: Comparison of advanced acceptance score with it’s variants . We compare the score selected as optimal by

A_{r}^{*}(\Delta)

and its variants: (i) individual measures and (ii) their different combinations. We assess on the basis of design criterion. The proposed

A_{r}^{*}(\Delta)

outperforms the other variants on all three datasets. The selected score are jointly optimal in terms of design criterion.

Fig. 3: Radar charts for comparing score selection of advanced acceptance score and it’s variants. Comparison in performed in terms of our design criterion: (i) rank deviation, (ii) relevance, (iii) trend deviation, and (iv) entanglement. To clearly discriminate between the relevance scores, we raised them as an exponent: $2^{\lambda\mathcal{R}}$ . In comparison to it’s variants, the advanced acceptance score satisfies all the criteria. This high end capabilities can be accredited to multi-faceted evaluation in the proposed. Exact numerical values can be inferred from Table I , while we illustrate a qualitative comparison in Fig. 4 .

Fig. 4: Qualitative comparison between advanced acceptance score and its variants over score/model selection (on Soli dataset). We conduct this analysis using: (i) histogram comparing selected DGBQA scores and ground truth (see the first row) and (ii) Gram matrix (to quantify entanglement). The score/model selected from $A_{r}^{*}(\Delta)$ (column (f)) has DGBQA scores comparable to the ground truth. While it attains the least entanglement (we require lighter values in off diagonal elements of the off diagonal blocks). This is also observed wrt [ 6 ] .

Fig. 5: Bars and whiskers plot to compare distribution of evaluation measures. The plot highlights that individual measures covers different range. This suggests that they convey different information. $A_{r}^{*}(\Delta)$ fuses them to generate a broader range of values. This is pivotal in differentiating between good and low quality scores.

	Soli					HandLogin					TinyRadar
Evaluation Measure	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$	Model	$\hat{r}$	$\mathcal{R}$	$\Psi$	$C_{d}$
RMSE	TPN (1.5,1.0)	1.00	8.62	1.39	0.331	MViT (1.0,1.0)	0.50	3.95	0.82	0.372	ViViT (1.0,2.5)	1.45	8.50	2.13	0.540
$cos$	TPN (1.5,1.0)	1.00	8.62	1.39	0.331	MViT (1.0,1.0)	0.50	3.95	0.82	0.372	ViViT (1.0,2.5)	1.45	8.50	2.13	0.540
DCG [ 16 ]	ViViT (1.0,1.0)	1.00	8.40	1.65	0.266	MViT (0.5,0.5)	0.50	3.93	3.19	0.313	ViViT (0.5,1.0)	1.45	8.44	2.26	0.502
$\tau$	ViViT (1.5,1.0)	2.09	8.05	1.76	0.315	TAM (1.0,2.5)	1.0	3.47	3.60	0.091	ViViT (0.5,1.0)	1.45	8.44	2.26	0.502
ERR [ 10 ]	MF (1.0,1.0)	1.00	8.64	1.78	0.268	MViT (0.5,0.5)	0.50	3.93	3.19	0.313	ViViT (0.5,1.0)	1.45	8.44	2.26	0.502

TABLE II: Comparison of the proposed with SOTA in terms of score selection. To conduct this study, we adapted retrieval quality measures [ 16 , 10 , 11 , 12 , 30 , 17 , 34 ] for our problem. Unlike existing measures, the score selected by

A_{r}^{*}(\Delta)

considers entanglement and trend deviation. This allows it jointly satisfy all the design criteria. For fair comparison, we also our quality measure: relevance (

\mathcal{R}

) in comparison. It is evident from the results, that

\mathcal{R}

consistently outperformed the other measures on the three datasets over rank deviation, relevance, and trend deviation. This can be accredited to its ability to reward both high and low ranking gestures as per their score values. A pictorial representation of this analysis has been illustrated with radar charts in Fig. 1 .

Fig. 6: Correlation analysis between the advanced acceptance score and the existing measures. The figure shows that the proposed measure correlates well with existing measures. This proves that it is both reliable and robust.

Fig. 7: Impact of varying scaling factors on $nA_{r}^{*}(\Delta)$ . We consider the optimal DGBQA for each of the dataset. Then, we plot $nA_{r}^{*}(\Delta)$ for different values of scaling factors. We vary a single factor at once. We observe that increasing or decreasing any of these factors results in significant changes in the $nA_{r}^{*}(\Delta)$ . Thus, this validates: (i) advanced acceptance score is sensitive to each of the individual parameters, and (ii) scaling factors can be utilized to select DGBQA scores as per user’s preference.

Fig. 8: Variation in surface plots of $nA_{r}^{*}(\Delta)$ vs $\lambda_{ID}~\&~\lambda_{ICGD}$ with changes in scaling factors. Here, the surface represents $nA_{r}^{*}(\Delta)$ values of MViT models, trained on Soli using different $\lambda_{ID}~\&~\lambda_{ICGD}$ . We study the changes in surface with varuation in scaling factors. The figure shows that increasing or decreasing any of these factors results in significant changes in the $nA_{r}^{*}(\Delta)$ (as also observed in Fig. 7 ). Further, when we compare across two surfaces across rows, we notice the entire surface getting raised or lowered. This suggests that individual evaluation measures can alter the score selection.

Biometric Identification

Biometric identification matches or verifies people and animals using distinctive physical, physiological, or behavioral signals.

See all 128 papers

Oct 23, 2026

VERA-Hand: A Verified Multidimensional Benchmark for 2D Hand Pose Estimation in Third-Person Dynamic Gesture Videos

Hand landmarks serve as input for gesture recognition, sign language understanding, and touchless human-computer interaction. Yet hand pose estimators are typically evaluated with frame-wise localization metrics such as PCK or MPJPE, which are computed over matched predictions and overlook missed or spurious hands, handedness errors, and temporal instability. To address these limitations, we present VERA-Hand, a benchmark that evaluates estimators using manually verified, frame-level 2D annotations for 10,690 hands across four complementary dimensions (detection, localization, identification, and temporal stability), as well as a joint Pose-F1 metric that requires both correct detection and accurate localization. Evaluating six estimators spanning hand-specific, whole-body, reconstruction-based, and foundation-model paradigms, we find that no estimator dominates: rankings reorder sharply across dimensions. VERA-Hand also exposes failures that localization-only and downstream evaluations may not reveal, including cross-hand overlap in over 85% of frames for a recent foundation model. A post-hoc cleanup case study demonstrates how the benchmark can validate corrective interventions, while stratified analyses show that robustness varies across motion-blur classes and hand regions. We release the verified annotations, protocol, and code at https://gibranbenitez.github.io/VERA_Hand/

Gibran Benitez-Garcia, Hiroki Takahashi

Graduate School of Informatics and Engineering, The University of Electro-Communications, Tokyo, Japan · Artificial Intelligence eXploration Research Center (AIX), The University of Electro-Communications, Tokyo, Japan · Meta-Networking Research Center (MEET), The University of Electro-Communications, Tokyo, Japan

Human Pose Estimation Object Detection

Ask PDF

Oct 8, 2026

Perceptually Grounded and Semantics-Aware Evaluation for Holistic Co-Speech Gesture Generation

Holistic and semantics-aware co-speech gesture generation has advanced rapidly, yet evaluation remains behind: objective metrics do not consistently reflect human perception, and semantic appropriateness remains difficult to quantify. We present a perceptually grounded and semantics-aware benchmark that combines standardized model comparison, human-centered metric validation, and fine-grained semantic evaluation. We first curate a list of 13 objective metrics covering different aspects, including distributional similarity, geometric fidelity, kinematic quality, cross-modal synchrony, and semantic appropriateness. For the semantic-appropriateness category, we propose a new metric, Semantic Gesture Preservation (SGP), which measures how far semantic gestures in the ground truth are preserved in the generated gestures. For this, we augment the BEAT2 dataset's annotations using a multi-modal LLM. We then conduct a perceptual study where 101 participants score generated gestures among five dimensions, including human-likeness, motion diversity, absence of animation errors, speech timing and content match. We systematically analyze objective metric--subjective score correlations. Unlike Semantic Score (SC), which shows no significant association with the evaluated perceptual dimensions, SGP is selectively aligned with speech-aware human judgments. We construct five target-specific composite metrics aligned with the subjective dimensions. These composites improve perceptual alignment across all five dimensions, with the largest gains for absence of animation errors and content match, indicating that complementary objective signals can better approximate human judgments than individual metrics alone. Overall, our results show that objective metrics require validation against subjective evaluations.

Nick Milkin, Lanmiao Liu, Esam Ghaleb +2

Max Planck Institute for Psycholinguistics, Nijmegen, The Netherlands · Utrecht University, Utrecht, The Netherlands · Donders Institute for Brain, Cognition and Behaviour, Nijmegen, The Netherlands

Co-Speech Gesture Generation Generative AI Evaluation

Ask PDF

Aug 13, 2026

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark

Humanoid motion tracking is central to teleoperation and whole-body imitation, yet evaluation often disagrees with what people perceive in videos. Kinematic errors average per-frame pose differences but miss the physical artifacts that matter most, particularly unstable support and incorrect contacts such as foot skating and mistimed touch-downs. Meanwhile, widely used test suites are small and lack the diversity needed to stress contact-rich, long-horizon behaviors. We introduce HumanTracker to make humanoid tracking evaluation both perceptually aligned and scalable. The HumanTracker benchmark contains approximately 153 hours of optical motion trajectories from multiple professional performers, organized into four motion families with text labels for fine-grained diagnosis. We further propose HumanScore, a preference-aligned metric trained on 12K motion pairs containing 24K motions. Across representative state-of-the-art trackers, HumanScore better predicts human preferences and reveals contact and stability failures that kinematic metrics often miss.

Dairu Liu, Zekun Qi, Jiayu Zeng +11

1Nankai University · 3Galbot · 2Tsinghua University +3

Human-in-the-Loop Evaluation Humanoid Motion Tracking Preference Learning Human Motion Understanding

Ask PDF

Abstract

Biometric Identification

Explore similar work

VERA-Hand: A Verified Multidimensional Benchmark for 2D Hand Pose Estimation in Third-Person Dynamic Gesture Videos

Perceptually Grounded and Semantics-Aware Evaluation for Holistic Co-Speech Gesture Generation

HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark