Learning faces with the BIAS model: On the importance of the sizes and locations of fixation regions

Predrag Neskovic, Ian Sherman, Liang Wu, Leon N. CooperNeurocomputing 2009 · DOI

TL;DR

This paper investigates how the size and location of "fixation regions" (areas of interest) impact a biologically inspired model's ability to learn and recognize faces. They found that smaller fixation regions and those closer to the center of the face lead to better performance, and importantly, the model is surprisingly robust to variations in these parameters.

What Problem Does This Paper Address?

Human visual perception is not uniform; we tend to focus our attention on specific, informative parts of an object, like the eyes, nose, and mouth of a face. This selective attention, combined with eye movements (saccades), allows us to learn and recognize objects efficiently. While many computer vision systems process images at a uniform resolution, biologically inspired models aim to mimic human visual processes. The authors' previous work introduced the Bayesian Integrate And Shift (BIAS) model, which learns object categories by integrating information from multiple localized regions (fixations). However, a critical question remained: how sensitive is this model to the precise size and location of these fixation regions? If the model is overly sensitive, its practical applicability would be limited. This paper addresses this by systematically exploring the impact of varying fixation region sizes and locations on the BIAS model's face recognition performance.

Methods

The core of the study uses the Bayesian Integrate And Shift (BIAS) model, a biologically inspired system that mimics human foveal vision and saccade-like movements. The model processes images by integrating information from a grid of receptive fields (RFs). A key component is the central receptive field (CRF), which is always positioned at the fixation point.

The paper focuses on two main aspects:

  • Feature Extraction: They use Gabor filters, inspired by simple cells in the visual cortex, to extract features from image regions.
  • Modeling Likelihoods: The model uses Gaussian distributions to represent the probability of feature detector outputs given a hypothesis about the fixation point's location.
  • Decision Rule: A likelihood ratio is computed to decide if a point belongs to a specific fixation region or the background.

The novelty lies in how they systematically vary the sizes and locations of these predefined fixation regions to test the BIAS model's robustness and performance. They compare performance using both "global" (across all images) and "local" (per image) thresholds for decision-making.

Datasets

The experiments were conducted using the 101 Caltech dataset, which contains 435 grayscale images of faces from approximately 30 subjects. The system was trained on background images and then on specific face regions. For testing, a subset of images was used, with a significant portion of faces differing from those in the training set to assess generalization.

Key Insights and Ideas

  • Robustness to Variation: The BIAS model demonstrates remarkable robustness to significant changes in the size and location of fixation regions. This is crucial for practical applications, as precise manual annotation of these regions can be challenging.
  • Size Matters (Inversely): Performance is inversely proportional to the size of the fixation region. Smaller regions lead to better recognition accuracy. This suggests that more focused attention on specific, informative parts of the face is beneficial.
  • Location, Location, Location: Fixation regions closer to the center of the face tend to yield better results. This aligns with the idea that central facial features (eyes, nose, mouth) are highly informative.
  • Teacher vs. Optimal: The regions selected by a human teacher are not necessarily the optimal ones for the model. The model can perform better with regions that are smaller or positioned differently than what a human might intuitively choose.
  • Internal Structure is Less Critical: The internal texture or structure of a fixation region (e.g., uniform vs. containing edges) has less impact on performance compared to its size and location.

Key Results

  • The paper demonstrates that the BIAS model can learn face categories with a sufficient number of training examples, even when fixation regions are varied significantly.
  • Performance was shown to be inversely proportional to the size of the fixation region. For instance, reducing the size by a factor of 10 (scale factor 0.1) generally improved performance compared to the teacher-defined regions.
  • Regions closer to the center of the face (e.g., "part 3" in their experiments) often achieved higher classification rates.
  • The system showed good performance even when fixation regions included areas outside the face, highlighting its robustness.
  • The use of a local threshold generally resulted in better performance than a global threshold, indicating that image-specific characteristics can influence optimal decision boundaries.

Limitations and Future Work

  • Teacher Dependence: While the model is robust to variations in fixation regions, the initial configuration and relative positioning of these regions are still influenced by the human teacher. Future work could explore more automated or adaptive methods for defining these regions.
  • Number of Fixations: The paper briefly touches upon the number of fixations per region, suggesting that larger regions might require more fixations. Further investigation into the optimal number of fixations for different region sizes and complexities could be beneficial.
  • Generalization with Small Datasets: While the model performs well with sufficient training data, the paper notes that with very few training images, a large number of fixations might lead to overfitting. Exploring strategies to improve generalization with limited data remains an area for future research.
  • Beyond Faces: The current study focuses on face recognition. Extending the investigation to other object categories would be valuable to understand the generalizability of these findings.

Summary updated

BibTeX
@inproceedings{neskovic2009learning,
  author = {Predrag Neskovic and Ian Sherman and Liang Wu and Leon N. Cooper},
  title = {{Learning faces with the BIAS model: On the importance of the sizes and locations of fixation regions}},
  booktitle = {Neurocomputing},
  year = {2009},
  doi = {10.1016/j.neucom.2008.09.018},
}
Figure