GaitFace: A Multimodal Dataset for Long-Range Person Identification

Alain Komaty¹, Luis Luevano¹, Vidit Vidit¹, Anjith George¹, Zeina Al Amine¹, Sébastien Marcel¹

¹Idiap Research Institute, Martigny, Switzerland

GaitFace Dataset

camera setup and facial resolution comparison

Top: visual comparison of facial resolution across the different cameras and heights used during acquisition. Bottom: sample frames from the "Walking Towards Camera" (WTC) scenario, showing the progression from long-range face and gait capture (~100m) to near-range facial capture (~3m).

Abstract

Efficient border control is becoming a significant global challenge, mainly due to severe congestion and extended passenger waiting times. To mitigate these bottlenecks and facilitate passenger flow, biometric technologies are increasingly deployed to streamline identity verification and enhance crossing efficiency. Technical limitations frequently impede biometric identification, particularly in long-range surveillance, where systems must deal with adverse atmospheric conditions and degraded image quality. While high-quality frameworks like BRIAR exist, they are frequently restricted to specific government agencies. This paper introduces GaitFace, a new public dataset that contains face and gait data captured at long distances. To ensure that the research reflects authentic border scenarios, we use Pre-Enrollment data, where a traveler registers via a mobile device, and "In-the-Wild" captures, which records individuals at a distance across multiple viewing angles and different cameras. Benchmarking SOTA face and gait models reveals that current architectures fail under low-resolution and elevated viewpoints despite success with optical assistance. GaitFace exposes these critical vulnerabilities, providing a rigorous public benchmark to drive more robust, unconstrained biometric research.

The Dataset

GaitFace was ethically and legally acquired from 70 consenting participants across two acquisition sessions separated by several weeks, amounting to approximately 2.7 TB of multimodal data: high-resolution RGB facial images in RAW format and multi-view gait recordings, captured at distances of up to 100 meters from both a ground-level and an elevated (~9–10 m) viewpoint. Participants were recorded outdoors under natural illumination and weather variations, following four standardized scenarios: normal walking, walking while simulating a phone call, walking with an accessory (bag or backpack), and walking with a different jacket.

map of recording stations and walking path — Map showing positions of the recording teams and the predefined walking path shared across scenarios and sessions.

Pre-enrollment face captures. Face enrollment was performed indoors with two mobile devices (iPhone 12 and Samsung Galaxy S9): each participant recorded a 15-second facial sequence including static and controlled head movements, mimicking a real-world remote pre-enrollment process. Outdoor face captures include high-quality optical-zoom images stored in RAW format (from the 3rd floor, ~100m), and native low-resolution face video from ground-floor and 3rd-floor cameras. At 100m without optical assistance the average face resolution is 18×22 pixels (IPD 4±1.2 px), versus 210×250 pixels (IPD 55±3.5 px) with optical zoom.

Face data samples, left to right: (a) Enroll, (b) Probe HQ (optical zoom), (c) 100m 3^rd Floor (3F), (d) 100m Ground Floor (GF), (e) 3m GF.

Outdoor gait captures. Gait was recorded simultaneously from ground-level cameras and an elevated (3rd-floor) viewpoint along the predefined trajectory. The ground-level camera position was intentionally varied between the two sessions to introduce additional viewpoint variation. Silhouettes are provided alongside the RGB frames.

gait multi-view samples — Walking direction variations (views V1–V4), RGB and silhouettes.

gait scenario variations — Walking style and appearance variations: Normal, Phone, Bag, Clothes.

Metadata. The dataset includes participant metadata (age, gender, and skin color) to enable benchmarking for algorithmic fairness and demographic robustness. The distribution is balanced in age and gender (45.7% female, 54.3% male).

demographic distributions — Demographic distributions of the dataset for age and skin color (Fitzpatrick scale).

Benchmarking: Long-Distance Face Recognition

We benchmark four SOTA face recognition models spanning a computational spectrum: EdgeFace (18.22M params), AdaFace IR50/IR101 (43–65M), and LVFace (140M+). Operational thresholds were calibrated on IJB-C.

FNMR (%) on the HQ (optical zoom) protocols. Thresholds calibrated on IJB-C at FMR = 1% and 0.1%.
Model	HQ-1		HQ-3		HQ-10		HQ-20
Model	1%	0.1%	1%	0.1%	1%	0.1%	1%	0.1%
IR50	1.3	1.8	1.1	1.6	0.9	1.2	0.9	1.1
IR101	0.9	1.3	0.9	1.1	0.8	1.0	0.8	0.9
EdgeFace	1.8	2.5	1.5	2.1	1.3	1.8	1.3	1.8
LVFace	0.9	1.2	0.5	0.7	0.4	0.6	0.4	0.6

With optical zoom (HQ protocols), all four systems achieve very good recognition results. The picture changes drastically without optical assistance.

GF protocol trend lines — **Protocol GF (100m, ground floor, no zoom):** TMR at a pre-calibrated threshold (FMR=1% on IJB-C) as the number of fused low-quality probe frames increases from 1 to 40. Even the best configuration only approaches a TMR of ~80%.

3F protocol ROC curves — **Protocol 3F (100m, elevated viewpoint):** ROC curves for the four systems collapse toward the chance line (AUC ≈ 0.5) — EERs range from 36.7% (IR101) to 43.9% (LVFace). Standard architectures are practically ineffective at this range.

TMR vs distance — **Protocol WTC:** TMR degradation as a function of distance across three FMR thresholds. All systems are near-perfect at close range, but degrade at distance; stricter thresholds trigger an earlier collapse.

Benchmarking: Long-Distance Gait Recognition

We evaluate silhouette-based methods — GaitBase (88M, trained on GREW), DeepGaitV2 (95M, GREW), GaitGL (3.1M, CASIA-B) — and the RGB-based BiggerGait (140M, CCPG), using the OpenGait framework with publicly available weights. Probes: Normal, Phone, Bag, Clothes; Gallery: Normal.

Mean Rank-1 accuracy (%) over the four camera views on the G1 and G2 gait protocols.
Method	Train	Rank-1 on G1				Rank-1 on G2
		Normal	Phone	Bag	Clothes	Normal	Phone	Bag	Clothes
GaitBase	GREW	26.80	21.89	23.33	24.03	21.02	18.76	19.95	15.86
DeepGaitV2	GREW	24.53	22.59	19.52	21.01	28.72	26.57	25.00	23.84
GaitGL	CASIA-B	12.25	12.50	11.07	12.08	8.46	7.05	7.07	7.37
BiggerGait	CCPG	8.78	7.67	9.64	10.67	3.18	3.00	3.55	3.86

Both silhouette- and RGB-based methods generalize poorly on GaitFace, highlighting the challenge of long-distance gait recognition. Performance is higher when the camera viewpoint shift between sessions is small (G1) and drops under covariate changes (phone, bag, clothing). GaitBase performs best under G1, DeepGaitV2 under the more challenging G2.

CMC curves G1 — CMC curves, protocol G1.

CMC curves G2 — CMC curves, protocol G2.

BibTeX


      @inproceedings{komaty2026gaitface,
        title={GaitFace: A Multimodal Dataset for Long-Range Person Identification},
        author={Alain Komaty and Luis Luevano and Vidit Vidit and Anjith George and Zeina Al Amine and S{\'e}bastien Marcel},
        booktitle={IEEE International Joint Conference on Biometrics (IJCB)},
        year={2026},
      }