TransFIRA: Transfer Learning for FIQA

• TransFIRA is a transfer learning framework for Face Image Recognizability Assessment that defines recognizability via class-center geometry in a deployed encoder's embedding space.
• It computes recognizability using metrics like Class-Center Similarity and Class-Center Angular Separation, enabling effective filtering and weighting during template aggregation.
• The approach achieves state-of-the-art performance on benchmarks like BRIAR and IJB-C by aligning predictions with the intrinsic discriminability of the encoder.

Searching arXiv for the TransFIRA paper and closely related FIQA work to ground the article in current literature. arXiv search query: "TransFIRA Transfer Learning for Face Image Recognizability Assessment" TransFIRA, short for Transfer Learning for Face Image Recognizability Assessment, is a framework for face image recognizability assessment (FIQA) that defines recognizability directly in the embedding space of a deployed encoder rather than through conventional visual quality proxies. It is designed for unconstrained face recognition settings such as surveillance, video, and web imagery, where pose, blur, illumination, and occlusion can cause conventional quality metrics to diverge from actual matchability. The framework introduces recognizability labels derived from class-center geometry, learns to predict those labels with a lightweight regression head attached to a pre-trained encoder, and uses the resulting scores for filtering and weighting during template aggregation. The paper presents TransFIRA as lightweight and annotation-free, with no external quality labels, heuristics, or backbone-specific recognition training, and reports state-of-the-art verification accuracy on BRIAR and IJB-C together with substantially improved correlation to true recognizability (Tu et al., 7 Oct 2025).

1. Definition, scope, and disambiguation

TransFIRA addresses the problem that, in unconstrained face recognition, raw visual “quality” is often a poor predictor of whether a face recognizer will actually match an image correctly. The framework defines recognizability as how confidently an embedding-based recognizer will assign the image to its correct identity, relative to other identities, in its own embedding geometry. In this formulation, a slightly blurred image can remain highly recognizable if its embedding lies close to the subject’s latent cluster and far from impostor clusters, whereas a visually clean image can have low recognizability if its embedding is near competing identities (Tu et al., 7 Oct 2025).

The framework is positioned against FIQA approaches that use handcrafted proxies such as pose, blur, illumination, and contrast; regress from images or embeddings to labels derived from external recognizers or human ratings; or rely on large generative or vision-LLMs. The stated limitations of those approaches are that they are often not encoder-specific, misaligned with decision geometry, or heavy and complex. TransFIRA is proposed as an alternative that derives recognizability labels from the geometry of a pre-trained encoder and then learns to predict them from images (Tu et al., 7 Oct 2025).

The term should also be distinguished from TransF, a translation-based knowledge graph embedding model introduced in “Knowledge Graph Embedding with Multiple Relation Projections” (Do et al., 2018). TransFIRA is not a variant of that model; it is a recognizability-assessment framework for face and body recognition (Tu et al., 7 Oct 2025, Do et al., 2018).

2. Geometric formulation of recognizability

The geometric core of TransFIRA begins with a labeled dataset

$\mathcal{D} = \{(x_i, y_i)\}_{i=1}^N,$

where $x_i \in \mathbb{R}^{H \times W \times 3}$ is an image and $y_i \in \{1,\dots,K\}$ is its identity label. A pre-trained encoder $\phi$ maps each image to an embedding

$z_i = \phi(x_i) \in \mathbb{R}^d.$

For each identity $j$, the class center is defined as the mean embedding for that identity. If a separate gallery set $\mathcal{G}$ is available,

$$\mu_j = \frac{1}{n_j^{\mathcal{G}}} \sum_{\substack{i \in \mathcal{G} \ y_i = j}} z_i,$$

and otherwise

$\mu_j = \frac{1}{n_j} \sum_{i : y_i = j} z_i.$

These centers anchor the recognizability definition in the deployed encoder’s own latent space (Tu et al., 7 Oct 2025).

TransFIRA defines Class-Center Similarity (CCS) as the cosine similarity between an embedding and the center of its own identity:

$$CCS_{x_i} = \frac{z_i^\top \mu_{y_i}}{\|z_i\|_2 \, \|\mu_{y_i}\|_2}.$$

This quantity measures how well aligned an image is with its class prototype. High CCS indicates that the sample is near the center of its identity cluster and is likely to be typical and reliable (Tu et al., 7 Oct 2025).

To capture confusion with competing identities, the framework defines Nearest Nonmatch Class-Center Similarity (NNCCS):

$x_i \in \mathbb{R}^{H \times W \times 3}$0

NNCCS identifies the most similar impostor center. High NNCCS indicates that some other identity lies dangerously close in embedding space (Tu et al., 7 Oct 2025).

The main margin quantity is Class-Center Angular Separation (CCAS):

$x_i \in \mathbb{R}^{H \times W \times 3}$1

CCS measures closeness to the correct center, NNCCS measures closeness to the nearest incorrect center, and CCAS is their margin. Under normalized embeddings and centers, cosine similarity reflects angular distance, so CCAS is proportional to the angular margin between the correct and impostor classes. This yields the paper’s central recognizability criterion: an image is recognizable when it lies closer to its own class center than to any other center, i.e.,

$x_i \in \mathbb{R}^{H \times W \times 3}$2

Large positive CCAS indicates high recognizability; values near zero indicate ambiguity near a decision boundary; negative values indicate misclassification or severe confusion (Tu et al., 7 Oct 2025).

The framework contrasts this geometry-preserving margin with prior ratio-style scores such as CR-FIQA’s Certainty Ratio,

$x_i \in \mathbb{R}^{H \times W \times 3}$3

arguing that the ratio distorts the margin by adding a constant and applying a non-linear transformation. In this view, the sign of CCAS has direct decision-boundary meaning, whereas the midpoint of the ratio does not (Tu et al., 7 Oct 2025).

3. Framework architecture and transfer-learning procedure

TransFIRA is presented as a three-stage framework. The first stage is geometric recognizability labeling: from labeled images and a pre-trained encoder, it computes class centers and then assigns each image two recognizability labels, CCS and CCAS. The second stage is transfer-learning recognizability prediction: a lightweight MLP head is attached to the pre-trained encoder and the model is fine-tuned end-to-end to regress CCS and CCAS from the input image. The third stage is recognizability-informed aggregation, which uses predicted CCAS for filtering and predicted CCS for weighting during template construction (Tu et al., 7 Oct 2025).

Formally, with encoder features $x_i \in \mathbb{R}^{H \times W \times 3}$4, the head $x_i \in \mathbb{R}^{H \times W \times 3}$5 predicts

$x_i \in \mathbb{R}^{H \times W \times 3}$6

with target

$x_i \in \mathbb{R}^{H \times W \times 3}$7

Training uses mean squared error,

$x_i \in \mathbb{R}^{H \times W \times 3}$8

Both the backbone $x_i \in \mathbb{R}^{H \times W \times 3}$9 and the head $y_i \in \{1,\dots,K\}$0 are fine-tuned end-to-end rather than training the head alone (Tu et al., 7 Oct 2025).

The supervision requirements are limited to identity labels. No FIQA labels, no human quality ratings, and no external IQA models are required. The paper states that TransFIRA can be applied to any encoder by computing embeddings on a labeled training set, computing class centers, deriving CCS and CCAS, attaching an MLP head, fine-tuning backbone and head with MSE, and then using predicted recognizability scores at inference time. The recognition function itself does not need to be retrained or modified; only the feature extractor is fine-tuned slightly for recognizability prediction (Tu et al., 7 Oct 2025).

Ablation results reported in the paper indicate that end-to-end fine-tuning is stronger than head-only training with a frozen backbone, and that random initialization of the backbone converges more slowly and performs worse. Starting from a recognition-pretrained backbone and fine-tuning jointly is reported to reach high correlation quickly, with convergence in approximately one epoch (Tu et al., 7 Oct 2025).

4. Recognizability-informed aggregation

Template-based verification aggregates multiple frames for one subject into a single template embedding. Standard practice is uniform averaging,

$y_i \in \{1,\dots,K\}$1

which treats all frames equally. TransFIRA replaces this with aggregation informed by the predicted recognizability scores (Tu et al., 7 Oct 2025).

The first step is filtering by predicted angular separation:

$y_i \in \{1,\dots,K\}$2

The cutoff is described as natural and parameter-free, because it is intrinsic to the encoder’s geometry and corresponds directly to retaining only frames predicted to lie on the correct side of the decision boundary (Tu et al., 7 Oct 2025).

The second step is weighting by predicted class-center similarity:

$y_i \in \{1,\dots,K\}$3

Here $y_i \in \{1,\dots,K\}$4 estimates how close the frame embedding lies to its class center. Weighting by CCS emphasizes compact, representative embeddings and downweights atypical frames even when they remain on the correct side of the boundary. The paper presents filtering and weighting as complementary operations: filtering removes harmful frames, whereas weighting emphasizes the most reliable frames among those retained (Tu et al., 7 Oct 2025).

This aggregation strategy is described as operating without external labels, heuristics, or backbone-specific training, and as not modifying the recognition loss or backbone architecture. A plausible implication is that TransFIRA is meant to function as a plug-in recognizability layer above an existing encoder, provided that an identity-labeled set is available for deriving geometry-based supervision (Tu et al., 7 Oct 2025).

5. Experimental results on faces

The paper evaluates TransFIRA on BRIAR Protocol 3.1, described as long-range surveillance imagery with challenging poses, blur, and occlusions, and on IJB-C, a standard unconstrained template-based verification benchmark. Two main backbones are used: CosFace (BRIAR), a Swin-B Transformer trained with CosFace loss on BRIAR, and ArcFace (WebFace), an iResNet-50 from InsightFace trained on WebFace12M (Tu et al., 7 Oct 2025).

On BRIAR Protocol 3.1, the reported template-level gains are large. With CosFace (Swin-B trained on BRIAR), uniform averaging yields TAR at FMR $y_i \in \{1,\dots,K\}$5 of approximately 0.75, the best prior FIQA baseline is approximately 0.81, and TransFIRA (Filter+Weight) reaches approximately 0.89. With ArcFace (WebFace-trained) evaluated on BRIAR, uniform averaging yields TAR@$y_i \in \{1,\dots,K\}$6 of approximately 0.08, baseline FIQA methods are reported as typically $y_i \in \{1,\dots,K\}$7, and TransFIRA (Filter+Weight) reaches approximately 0.60, described as more than 3× improvement over average (Tu et al., 7 Oct 2025).

On IJB-C, the reported gains depend on backbone. For CosFace (BRIAR), average aggregation gives TAR@$y_i \in \{1,\dots,K\}$8 of approximately 0.24, the best baseline is approximately 0.37 with CR-FIQA(L), and TransFIRA (Filter+Weight) reaches approximately 0.86. For ArcFace (WebFace), average aggregation gives approximately 0.79, CR-FIQA(L) gives approximately 0.87, and TransFIRA (Filter+Weight) gives approximately 0.87–0.88, effectively matching or slightly surpassing the strongest baseline at strict FMR (Tu et al., 7 Oct 2025).

At the image level, the paper evaluates recognizability prediction using Spearman correlation and FNMR–ERC curves. On BRIAR (CosFace), CCS predictions achieve Spearman correlation of approximately 0.86, whereas the best baseline is reported as $y_i \in \{1,\dots,K\}$9, with ERC AUCs roughly half those of the baselines. On IJB-C, TransFIRA is reported to have the highest or second-highest correlation and the lowest ERC AUC across most settings. These findings support the claim that the predicted scores track the encoder’s true discriminability more closely than prior FIQA methods (Tu et al., 7 Oct 2025).

The paper summarizes these results by stating that TransFIRA achieves state-of-the-art or near state-of-the-art template verification across BRIAR and IJB-C, and that its CCS/CCAS predictions nearly double the correlation with ground-truth recognizability on BRIAR relative to the best baseline. It also reports cross-dataset robustness: training recognizability on BRIAR and applying it to IJB-C, and vice versa, still yields competitive TAR, often close to in-domain tuned models. This suggests that CCS/CCAS-based recognizability captures a property of the encoder’s geometry that is relatively stable under dataset shift (Tu et al., 7 Oct 2025).

6. Extensions beyond faces, explainability, and limitations

TransFIRA is extended beyond face recognition to body recognition using the SemReID encoder. The same pipeline is used: class centers are computed in body-embedding space, CCS, NNCCS, and CCAS are derived, and a recognizability head is trained. The paper reports that body embeddings exhibited highly saturated cosine similarities, with CCS and NNCCS near 1.0, giving little variance. To address this, it introduces a sigmoid calibration that pulls CCS toward 1 and NNCCS toward 0, thereby spreading them across $\phi$0 so that CCAS regains useful variation. On BRIAR bodies, baseline averaging gives TAR@$\phi$1 of approximately 0.085, calibrated CCAS weighting gives approximately 0.104, and calibrated CCAS Filter+Weight gives approximately 0.108. The paper describes this as the first recognizability-aware body recognition assessment (Tu et al., 7 Oct 2025).

The framework also supports encoder-grounded explainability because CCS and CCAS are directly tied to the embedding space of the recognizer. The paper uses perturbation analysis to evaluate how recognizability changes under blur and other degradations. For Gaussian blur on IJB-C, light blur is reported to sometimes improve CCS and CCAS by removing distractors or minor misalignment, while moderate and heavy blur cause monotonic decline. Predicted scores track ground-truth CCS and CCAS under blur with Spearman correlation of approximately 0.62–0.66. This is used to argue that visual degradation and recognizability are not necessarily monotonic, and that encoder-grounded measures capture this non-trivial relationship better than appearance-only quality metrics (Tu et al., 7 Oct 2025).

The framework has several explicit assumptions and limitations. It requires identity-labeled training data to compute class centers and therefore is not fully unsupervised. CCS and CCAS are said to be most meaningful for embedding-based encoders with a cosine or normalized-inner-product decision rule. In regimes with extreme similarity saturation, such as the reported body-recognition case, calibration is required. The method also requires fine-tuning the full backbone for recognizability prediction, which may be problematic when the deployed model must remain frozen; the paper mentions future work on distilling recognizability into smaller student heads. Finally, CCS and CCAS are defined relative to existing class centers, so the training distribution must approximate the deployed class distribution well enough for the learned geometry to remain indicative (Tu et al., 7 Oct 2025).