Research RadarFace RecognitionarXivJuly 2026

Monthly arXiv Radar

July 2026 Face Recognition Papers: Synthetic Foundation Models, Drive-Through NIR, and Lensless Cameras

July's face recognition papers move the discussion from benchmark-only accuracy toward the data and capture systems that determine deployment quality. The selected work tests privacy-conscious synthetic training at two data scales, introduces a cross-spectral border-control dataset with through-glass probes, and measures what happens when the camera itself is replaced by a lensless optical encoder.

What This Month Signals

Synthetic-only adaptation can make a general vision foundation model competitive for face recognition, but the winning strategy changes sharply with data volume. DriveFace then shows that a strong embedding model still faces meaningful errors when visible enrollment images are matched to moving NIR probes through automotive glass. LFD pushes the capture stack further by treating lensless optics, reconstruction, detection, and recognition as linked components. The common lesson is that deployment performance depends on the whole acquisition-to-embedding pipeline, not only the recognition loss.

Paper 012026-07-27cs.CV

IJCB-AFMFR 2026: Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data

arXiv PDF

Authors & Institutions

Tahar Chettaoui

Fraunhofer Institute for Computer Graphics Research IGD, Germany

Guray Ozgur

Fraunhofer Institute for Computer Graphics Research IGD, Germany

Technical University of Darmstadt, Germany

Eduarda Caldeira

Fraunhofer Institute for Computer Graphics Research IGD, Germany

Technical University of Darmstadt, Germany

Arturas Nakvosas

Vilnius University, Lithuania

Hatef Otroshi Shahreza

Idiap Research Institute, Switzerland

Sébastien Marcel

Idiap Research Institute, Switzerland

Rishabh Shukla

Indian Institute of Technology Jammu, India

Aditya Takkar

Indian Institute of Technology Jammu, India

Rushil Khullar

Indian Institute of Technology Jammu, India

Lalak Yadav

Indian Institute of Technology Jammu, India

Gourav Gupta

ArogyaPandit Private Limited, India

Anant Gupta

ArogyaPandit Private Limited, India

Shiqi Yu

Southern University of Science and Technology, China

Vitomir Struc

University of Ljubljana, Slovenia

Naser Damer

Fraunhofer Institute for Computer Graphics Research IGD, Germany

Technical University of Darmstadt, Germany

Fadi Boutros

Fraunhofer Institute for Computer Graphics Research IGD, Germany

What Problem It Solves

Earlier synthetic-data challenges mixed together generator quality, backbone choice, and training recipes. This benchmark fixes CLIP ViT-L/14 and the IDPERTURB generator so that full fine-tuning, LoRA, rank-stabilized LoRA, and projection-only adaptation can be compared under the same identity data and evaluation suite.

Key Result

Full fine-tuning with Sub-Center ArcFace leads the Full Data Track at 95.51% average accuracy on the small benchmarks and 87.42% TAR at FAR 1e-5 on IJB-C, versus 76.71% for the FRoundation baseline. Under limited data, rsLoRA is more robust: Idiap-BSP reaches 94.52% average accuracy and 81.13% IJB-C TAR at FAR 1e-5. The fairness evaluation also changes by regime, with the full-data winner reaching 91.70% average RFW accuracy and the limited-data winner 88.92%.

Abstract

This paper presents a summary of the Competition on Adapting Foundation Models for Face Recognition Using Synthetic Training Data (AFMFR), held at the 2026 International Joint Conference on Biometrics (IJCB 2026). The competition received a total of eight valid submissions from four distinct teams across two complementary tracks: a Full Data Track, in which participants adapt the CLIP ViT-L/14 foundation model using large-scale synthetic identity data, and a Limited Data Track, designed to reflect more resource-constrained adaptation regimes. All training data was generated exclusively using IDPERTURB. Submitted solutions are ranked based on verification and identification performance across a diverse suite of benchmarks, including LFW, CFP-FP, AgeDB-30, CALFW, CPLFW, IJB-B, IJB-C, and TinyFace, using the Borda count method. Fairness evaluation is additionally conducted on the RFW dataset across four demographic groups. The results demonstrate that adaptation of the CLIP foundation model with synthetic training data substantially improves over the off-the-shelf model and, in several cases, surpasses the baseline. Notably, full fine-tuning with Sub-Center ArcFace (DMSTI-Neurotechnology) leads the Full Data Track, while rank-stabilized LoRA adaptation (Idiap-BSP) proves most effective under limited-data conditions.

Research Starting Point

Face recognition developers increasingly lack broad, consented real-face training sets, while general vision foundation models are not sufficiently discriminative at strict biometric operating points without adaptation. The competition asks whether synthetic identities can provide that adaptation and how much trainable capacity is appropriate when the available synthetic set changes by more than an order of magnitude.

Method

The competition defines a Full Data Track with about three million images from 35,000 synthetic identities and a Limited Data Track with 250,000 images from 5,000 identities. Eight valid submissions are scored with a Borda aggregation across five small verification sets, IJB-B, IJB-C, and TinyFace, while RFW measures performance dispersion across four demographic groups. The submitted methods explore full-backbone updates, LoRA and rsLoRA ranks, margin losses, augmentation, and lightweight projection tuning.

Paper Summary

For teams adapting a large visual backbone, synthetic data volume should drive the tuning strategy. Full fine-tuning pays off at multi-million-image scale, while high-rank rsLoRA acts as a useful regularizer when identities and images are constrained; a single recipe is unlikely to be optimal for both settings.

Paper 022026-07-15cs.CV

DriveFace: A Cross-Spectral Through-Glass Face Dataset for On-the-Move Vehicular Border Control

arXiv PDF

Authors & Institutions

Anjith George

Idiap Research Institute, Switzerland

Luis S. Luevano

Idiap Research Institute, Switzerland

Alain Komaty

Idiap Research Institute, Switzerland

Zeina Al Amine

Idiap Research Institute, Switzerland

Vidit Vidit

Idiap Research Institute, Switzerland

Sebastien Marcel

Idiap Research Institute, Switzerland

University of Lausanne, Switzerland

What Problem It Solves

DriveFace fills the absence of a public benchmark that connects smartphone RGB enrollment to external near-infrared vehicle capture. It also includes a presentation-attack subset so recognition and capture security can be studied under the same operational geometry.

Key Result

Outdoor matching is relatively strong, with AdaFace at 99.45% AUC, 2.69% EER, and 97.42% Rank-1. Simulated tint is harder: the best AUC is 96.23%, and xEdgeFace records 8.09% EER and 88.63% verification at 1% FAR. The PAD baselines also deteriorate on unseen attacks; the best reported ACER for unseen masks is 26.20%, despite low errors in the known-attack protocol.

Abstract

The continuous growth in cross-border mobility places increasing pressure on existing border control infrastructures, motivating on-the-move biometric authentication, in which travellers are identified directly inside their vehicles at checkpoints. Face recognition is well-suited to this setting, as it can be acquired passively and at a distance. Its development, however, is hindered by the lack of representative datasets: existing benchmarks are collected in controlled environments and do not capture the challenges inherent to vehicular acquisition, including motion blur, variable illumination, occlusions, and cross-spectral enrollment. To address this gap, we introduce a dataset for on-the-move face recognition in border-control scenarios, comprising NIR vehicle-crossing videos paired with smartphone-based pre-enrollment data. Baseline evaluations with state-of-the-art models show clear performance limitations under these realistic conditions, highlighting the need for dedicated methods to advance the field.

Research Starting Point

Vehicle checkpoints would benefit from identifying pre-enrolled travelers without requiring them to leave the car, but the probe is captured through reflective or tinted glass, often while the vehicle or subject is moving. Existing indoor VIS-NIR datasets do not combine that optical degradation with profile pose, weather, speed, and cross-session enrollment.

Method

The dataset follows 70 consenting volunteers over two sessions and records smartphone enrollment plus outdoor, indoor-car, and simulated-tint NIR probes. Metadata covers glass tint, illumination, head pose, weather, and vehicle speed. Identity-disjoint train and test protocols evaluate AdaFace, LVFace, EdgeFace, and a cross-spectrally adapted xEdgeFace using AUC, EER, verification rate at 1% FAR, and Rank-1 identification; separate PAD protocols cover known and unseen print or mask attacks.

Paper Summary

DriveFace is valuable as an acceptance-test dataset for vehicle and border deployments because it keeps enrollment, sensor spectrum, glass, motion, and spoofing in one protocol. Its baseline numbers warn that good unconstrained face recognition scores do not guarantee a comfortable security margin under tinted cross-spectral capture.

Paper 032026-07-11cs.CV

LFD: Enabling Real-World Lensless Face Recognition with a Large-Scale Dataset

arXiv PDF

Authors & Institutions

Junho Kim

Department of Electrical and Computer Engineering, Rice University, Houston, Texas, USA

Salman S. Khan

Department of Electrical and Computer Engineering, Rice University, Houston, Texas, USA

Sara Wan

Department of Electrical and Computer Engineering, Rice University, Houston, Texas, USA

Tomi Kuye

Department of Electrical and Computer Engineering, Rice University, Houston, Texas, USA

Ashok Veeraraghavan

Department of Electrical and Computer Engineering, Rice University, Houston, Texas, USA

What Problem It Solves

The field lacked a sizable dataset pairing real lensless measurements, reconstructed images, and conventional face images across indoor and outdoor conditions. That gap made it difficult to separate reconstruction quality from detector and recognizer failure or to test whether a model transfers between different optical encoders.

Key Result

On Prototype 1, the LFD-trained recognizer reaches AUC 0.946 on Indoor Easy, 0.808 outdoors, 0.851 on the complete set, and 0.778 on the hard set. The corresponding standard-image model scores 0.884, 0.625, 0.763, and 0.599. At 0.05% FPR in cross-modal matching, LFD reconstructions reach 77.30% TPR versus 66.02% for the earlier FCFD FlatNet data, and the advantage persists on a second optical prototype.

Abstract

Face recognition is a ubiquitously used computer vision task that has a wide range of applications ranging from everyday smartphone biometrics to high-stakes security systems. Most face recognition systems rely on traditional cameras, which often suffer from limitations such as bulky form factors, high costs, and limited privacy protection. To address these limitations, lensless cameras have emerged as an alternative. Lensless cameras use thin optical encoders, enabling smaller size, lower cost, and greater design flexibility. These cameras are typically paired with reconstruction algorithms that convert raw captures into recognizable images. However, reconstructed images often contain artifacts, and the reconstruction methods struggle to generalize well to real-world conditions. Furthermore, existing face datasets do not account for the artifacts present in lensless images. To address this issue, we introduce the Lensless Face Dataset (LFD). LFD comprises 21,080 lensless raw measurements, reconstructions, and standard images of faces captured under diverse lighting, angle, and distance. Our key contributions are: (1) Real-world lensless face data: LFD focuses on capturing a diverse face dataset with varying levels of artifacts introduced under different environments; (2) In-the-wild captures: 4,976 images are captured in outdoor settings with varying intensities of natural light and different background patterns; (3) Multiple lensless devices: LFD includes face images collected from three different types of lensless cameras, each with a unique optical encoder. We use this hardware diversity to demonstrate generalization across different lensless cameras. Through comprehensive evaluations and analysis, we show that LFD effectively captures shared features and artifacts across different lensless imaging devices, making it a valuable dataset for advancing lensless face recognition.

Research Starting Point

Lensless cameras replace bulky lenses with thin coded optical elements, which can reduce size and expose less human-readable imagery at the sensor. Their reconstructions contain device-specific blur, color distortion, and vignetting, however, so recognition models trained on normal photographs or simulated measurements do not transfer reliably to physical captures.

Method

LFD collects 21,080 raw lensless measurements, reconstructions, and standard images, including 4,976 outdoor images and captures from three lensless camera designs. The authors train reconstruction and face-detection components separately, then compare SENet recognizers trained on standard VGGFace2 images, simulated lensless reconstructions, or the physical LFD training split. Verification uses balanced positive and negative pairs across easy, outdoor, complete, and hard subsets, followed by cross-device tests.

Paper Summary

LFD shows that a lensless biometric product must train against real sensor artifacts rather than treating reconstruction as a transparent preprocessing step. It provides a concrete basis for evaluating compact or privacy-oriented cameras, while the remaining outdoor gap shows that the hardware benefit still carries a substantial recognition cost.