Ocular Recognition: Interoperability, Multi-Resolution, and Spectrum Challenges

Fabian Stockhardt, Juan E. Tapia, Christoph BuscharXiv 2026 · arXiv

TL;DR

This paper investigates how to build ocular recognition systems that work across the wide range of resolutions and light spectrums (RGB vs. NIR) encountered in "biometrics on the move" scenarios. The authors demonstrate that while specialized models excel at specific resolutions, a Shared-Trunk architecture trained on a diverse, full-resolution image pool offers the best balance for direct interoperability between different capture conditions.

What Problem Does This Paper Address?

Traditional ocular recognition (iris and periocular) relies on controlled environments—think of a stationary subject looking directly into a high-res Near-Infrared (NIR) sensor. However, real-world applications like border control require "biometrics on the move."

As a traveler approaches a gate, the camera captures a sequence of images where the ocular region grows from a tiny, low-resolution visible-light (VIS) crop to a detailed NIR iris image. This creates three major hurdles:

  1. Heterogeneity: How do you compare a 40x40 pixel RGB periocular crop to a 160x160 pixel NIR iris image?
  2. Data Imbalance: We have millions of low-quality eye crops from face datasets but very few high-quality, identity-labeled iris samples.
  3. Architectural Choice: Should we use resolution-specific "specialist" models, or one giant model that tries to do everything?

Methods

The researchers evaluated three architectural families based on the AdaFace framework, varying the degree of parameter sharing:

  • INDEPENDENT: Three separate specialists (trained for 40x40, 80x80, and 160x160 inputs) with no shared parameters.
  • IBSH (Independent Backbones with Shared Heads): Each resolution has its own backbone, but they share the final classification "heads" to force the different models to map identities into the same embedding space.
  • SHARED-TRUNK: Resolution-specific front-ends feed into a single, shared residual trunk and embedding projection. This model is the most efficient, reducing parameters by ~64% compared to the independent approach.

They also tested two training regimes:

  • Routed (1/3): Specialists only see data that naturally falls into their resolution bin.
  • Full (3/3): Every specialist/route is trained on the entire dataset, with images resized to fit the route.

Datasets

The study utilized a massive, heterogeneous training pool:

  • Visible Spectrum (VIS): Eye crops extracted from VGGFace2 (millions of samples).
  • Cross-Spectral (VIS/NIR): CASIA NIR-VIS 2.0 and ND-NIVL.
  • NIR Iris-Native: PolyU Cross-Spectral Iris, CASIA-Iris-Lamp, and CASIA-Iris-Thousand.

Evaluation was performed on five "unseen" datasets to test generalization: FERET, UFPR-Periocular, MCXFace, IIT Delhi Iris, and MMU1 Iris.

Key Insights and Ideas

  • The Specialization vs. Interoperability Trade-off: Models trained only on their specific resolution (Routed) become highly accurate specialists but fail miserably when asked to compare their templates against those from other resolutions.
  • Lower-Route Rerouting is a Tough Baseline: A common practical fix is to downsample high-res images to match low-res ones before comparison. The authors found this "common-route" approach is surprisingly robust and often outperforms more complex cross-resolution strategies.
  • More Data $\neq$ Uniform Improvement: Training on the "Full" image pool (3/3) significantly helped the high-resolution (160x160) route but actually weakened the performance of the low-resolution (40x40) specialists in some cases. This suggests that "more data" can introduce noise or artifacts (via upsampling) that distract specialists.
  • Shared-Trunk Efficiency: The Shared-Trunk model proved that you don't need unique weights for every resolution. By sharing the core feature extraction logic, the model achieves competitive accuracy while being much easier to deploy.

Key Results

  • Interoperability: The Shared-Trunk model trained on the full image pool achieved the lowest overall Equal Error Rate (EER) for same-resolution verification at 6.20%.
  • Cross-Resolution: For direct cross-resolution scoring (comparing different input sizes), the Shared-Trunk model significantly outperformed the IBSH approach, reducing EER from ~17% to ~7.5% in the all-condition macro.
  • Parameter Savings: The Shared-Trunk architecture requires only 18.58M parameters for the library, compared to 52.68M for the Independent/IBSH versions—a 64.73% reduction in storage requirements without sacrificing performance.

Limitations and Future Work

  • Confounding Factors: In the current datasets, resolution often co-varies with the light spectrum (low res is usually VIS, high res is usually NIR). It is difficult to isolate whether the model is specializing in "resolution" or "spectrum."
  • Identity Overlap: The authors note that while datasets are separate, they cannot guarantee that the same individual doesn't appear across different internet-scraped training sets, which could slightly bias results.
  • Future Directions: The authors suggest exploring more sophisticated "synchronized separate-head" controls and investigating how temporal aggregation (combining multiple frames as a subject moves) could further boost accuracy.

Summary updated

Abstract

Ocular recognition is typically developed under controlled acquisition conditions, limiting its use in semi-controlled and on-the-move scenarios. A key obstacle is the lack of representative data spanning real-world capture conditions. At a border checkpoint, for example, images may range from low-resolution visible-light periocular crops (RGB) to high-resolution near-infrared iris images (NIR). Using this heterogeneous evidence raises three questions: whether resolution-specific processing improves recognition, how to compare templates from different processing routes, and how to allocate training data across resolution specialists. We investigate resolution-specific processing and parameter sharing using three AdaFace-based architecture families: fixed-resolution specialists; Independent Backbones with Shared Heads (IBSH); and models with resolution-specific front-ends for 40x40, 80x80, and 160x160 inputs, followed by a shared identity trunk and embedding projection. We evaluate each architecture with route-specific training subsets and with a full-pool strategy, in which every specialist uses the complete selected training set. Results reveal a trade-off among same-resolution specialization, cross-resolution interoperability, and common-route re-extraction. Route-specific training produces distinct performance profiles across routes, whereas full-pool training makes route performance more uniform and primarily benefits the 160x160 route. Thus, adding training data does not improve all resolution routes equally.
BibTeX
@misc{stockhardt2026ocular,
  author = {Fabian Stockhardt and Juan E. Tapia and Christoph Busch},
  title = {{Ocular Recognition: Interoperability, Multi-Resolution, and Spectrum Challenges}},
  year = {2026},
  eprint = {2611.20219},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.20219},
}
Figure