TransFIRA+: Transfer Learning for Biometric Recognizability Assessment

Allen Tu, Kartik Narayan, Joshua Gleason, Jennifer Xu, Matthew Meyn, Tom Goldstein, Vishal M. PatelarXiv 2026 · arXiv

TL;DR

TransFIRA+ shifts biometric quality assessment from "does this image look good to a human?" to "is this image recognizable to my specific AI model?" By deriving recognizability labels directly from a pretrained encoder's embedding space and introducing a novel sigmoid calibration for multimodal fusion, it achieves a massive performance leap in long-range surveillance, raising face verification accuracy (TAR at $10^{-6}$ FMR) from 0.085 to 0.598.

What Problem Does This Paper Address?

In real-world surveillance, biometric systems face a barrage of "degradations": extreme distances, motion blur, poor lighting, and awkward poses. Traditionally, we’ve used Face Image Quality Assessment (FIQA) to filter out bad frames. However, traditional FIQA relies on visual heuristics (like sharpness) or human-curated labels.

The problem is that visual quality $\neq$ recognizability. A blurry face might still be perfectly discriminative to a deep neural network, while a clear face at a weird angle might cause a total failure. Existing methods are often "detached" from the actual decision-making logic (the decision geometry) of the recognition model being used. TransFIRA+ aims to ground quality assessment directly in the encoder's own embedding space.

Methods

The authors introduce a lightweight framework called TransFIRA (Transfer Learning for Face Image Recognizability Assessment) and its multimodal extension, TransFIRA+.

1. Encoder-Grounded Labels

Instead of human labels, the system uses two geometric metrics derived from a pretrained, fixed encoder:

  • Class-Center Similarity (CCS): How well an image embedding aligns with the average embedding (center) of its true identity.
  • Class-Center Angular Separation (CCAS): The margin between the similarity to the correct identity and the similarity to the "nearest" impostor identity.

As shown in the architecture diagram below, a "Recognizability Prediction Head" is trained to predict these values directly from raw images.

The TransFIRA architecture: A pretrained encoder's embedding space is used to generate CCS and CCAS labels, which a secondary head learns to predict from raw pixels.

2. Principled Filtering and Weighting

  • Filtering: If predicted $CCAS \leq 0$, the model essentially predicts that the image is more likely to be an impostor than the correct person. This provides a "natural," parameter-free cutoff for discarding junk frames.
  • Weighting: During "template aggregation" (combining multiple video frames into one representation), frames with higher predicted CCS are given more weight.

3. Sigmoid Calibration (The "Plus" in TransFIRA+)

When moving to whole-body recognition, the authors found that scores often "saturate"—meaning almost all frames get very high similarity scores (e.g., 0.97 to 0.99). This makes direct weighting useless because every frame looks equally "good." They introduce a post-hoc sigmoid calibration to spread these saturated scores across a 0-1 range, restoring the ability to distinguish between high- and low-utility frames.

Sigmoid calibration spreads saturated body recognition scores, transforming them from effectively uniform weights into informative ones.

Datasets

  • BRIAR: A challenging long-range surveillance dataset (faces and bodies at distances up to 1000m).
  • IJB-C: A standard benchmark for unconstrained face recognition.
  • VideoLINCS (DS0001 & VL-MEVID): Two unseen multi-camera surveillance datasets used to test how well the model generalizes to new environments without retraining.

Key Insights and Ideas

  • Recognizability is Model-Specific: An image might be "high quality" for one model but "low quality" for another if their embedding geometries differ. TransFIRA captures this model-specific behavior.
  • The CCAS Cutoff: The $CCAS > 0$ threshold is a mathematically grounded way to say "this frame is more likely to help than hurt."
  • Calibration is Essential for Fusion: In multimodal settings (face + body), if one modality's quality scores are saturated, it can't effectively "step back" when it's uncertain, leading to poor fusion. Calibration fixes this.
  • Zero-Shot Transfer: A recognizability predictor trained on one dataset (BRIAR) generalizes remarkably well to entirely different camera setups and sites, outperforming standard detector-confidence metrics.

Key Results

The quantitative improvements are most dramatic on the difficult BRIAR surveillance benchmark:

  • Face Verification: Using TransFIRA's filtering and weighting raised the True Accept Rate (TAR) at a strict False Match Rate (FMR) of $10^{-6}$ from 0.085 (baseline) to 0.598 for an ArcFace model.
  • Body Recognition: Calibrated weighting improved body-only identification mAP on the unseen DS0001 dataset by over 3 points compared to uniform averaging.
  • Multimodal Fusion: On the VL-MEVID dataset, calibrated TransFIRA+ weights improved fused face/body identification Rank-1 accuracy by 2.9% over standard detector-confidence weighting.

Detailed comparisons against other state-of-the-art FIQA methods (like MagFace and CR-FIQA) consistently show TransFIRA+ taking the lead, especially in the most challenging low-FMR regimes.

Table showing TransFIRA+ outperforming a wide array of baseline FIQA methods on the BRIAR and IJB-C benchmarks.

Limitations and Future Work

  • Backbone Dependency: The recognizability head is tied to the specific recognition encoder. If you upgrade your face matcher, you need to retrain the (admittedly lightweight) TransFIRA head.
  • Tracking Integration: Currently, the system assesses frames independently. Future work could integrate this with temporal tracking to select optimal "glimpses" or gait cycles.
  • Generative Models: The authors suggest that semantic or prompt-conditioned prototypes could eventually adapt this framework to vision-language models.

Summary updated

Abstract

Biometric recognition in surveillance, video, and web imagery must contend with extreme variation in pose, blur, illumination, occlusion, and distance, where visual quality metrics fail to predict whether inputs are recognizable to the deployed encoder. Existing face image quality assessment methods rely on visual heuristics, curated annotations, or costly generative pipelines, leaving their predictions detached from the encoder's decision geometry. We introduce TransFIRA, a lightweight framework that grounds recognizability directly in embedding space. TransFIRA derives class-center similarity (CCS) and class-center angular separation (CCAS) from a pretrained encoder's embeddings and learns to predict them from images without quality annotations. Using predicted CCS for template weighting and CCAS>0 as a natural filtering threshold yields state-of-the-art face verification performance on BRIAR, raising TAR at $10^{-6}$ FMR from 0.085 with uniform averaging to 0.598. The predicted scores also reveal how image degradations affect encoder-specific recognizability. TransFIRA+ extends this framework to cross-camera multimodal person identification in untrimmed surveillance video. In this setting, body recognizability scores saturate within a narrow range, making direct weighting effectively uniform. We introduce a post-hoc sigmoid calibration that transforms these saturated scores into informative weights without retraining. After calibration, the BRIAR-trained TransFIRA predictor improves fused face and body identification on two unseen multi-camera datasets, outperforming detector-confidence weighting and uncalibrated CCS-rank weighting. Together, TransFIRA and TransFIRA+ demonstrate that encoder-grounded recognizability generalizes across biometric modalities and recognition settings.
BibTeX
@misc{tu2026transfira,
  author = {Allen Tu and Kartik Narayan and Joshua Gleason and Jennifer Xu and Matthew Meyn and Tom Goldstein and Vishal M. Patel},
  title = {{TransFIRA+: Transfer Learning for Biometric Recognizability Assessment}},
  year = {2026},
  eprint = {2611.23442},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.23442},
}
Figure