If you want to build a computer vision model that predicts where humans look in an image, a remarkably effective baseline is to simply output a blurry circle directly in the middle of the frame.
This phenomenon is known as center bias. It stems from two distinct sources: a biological tendency for human eye movements to naturally drift toward the center of a visual field (Wloka et al. [1]), and a compositional "photographer's bias" where creators naturally place primary subjects of interest in the center of a photograph (Schauerte et al. [2]). Even graph-based computational architectures inherently produce center-weighted maps because central nodes have a shorter average distance to all other nodes in the network (Harel et al. [3]).
While center bias reflects real-world human behavior, it introduces a severe confounding variable in computer vision. It makes it incredibly difficult to evaluate whether a model has actually learned to understand visual features, or if it is simply exploiting a global spatial shortcut.
The Confounding Problem in Evaluation
When evaluating a saliency model using standard metrics like the Area Under the ROC Curve (AUC), the model's predictions are treated as a binary classifier to distinguish human fixation points (positives) from random background pixels (negatives).
Because human fixations are heavily clustered in the center, a completely "blind" model that predicts a static 2D Gaussian blob will achieve exceptionally high performance scores (Borji et al. [4]). This mathematical quirk masks model deficiencies. As Kuemmerer et al. demonstrate, failing to decouple this baseline makes it impossible to isolate the actual information gain provided by a model's feature extraction [5].
This confounding effect extends beyond pure saliency prediction. In Visual Question Answering (VQA) and action recognition, task-independent center biases can artificially inflate the perceived performance of attention mechanisms, making it seem as though a model is performing complex semantic reasoning when it is merely relying on generic photographic composition (Das et al. [6]).
Decoupling Bias via Shuffled Evaluation
To accurately measure a model's ability to extract stimulus-driven features, researchers developed center bias decoupling techniques, most notably Shuffled AUC (sAUC).
The core intuition behind Shuffled AUC is a clever modification of the negative distribution. Instead of sampling negative background pixels uniformly from the same image, the negative distribution is drawn from the actual human fixations on completely different images within the same dataset.
Standard AUC:
[True Fixations (Center-Heavy)] vs. [Uniform Random Pixels (Spread Out)]
--> A central blob model easily separates these, scoring artificially high.
Shuffled AUC:
[True Fixations (Center-Heavy)] vs. [Other Images' Fixations (Center-Heavy)]
--> A central blob model cannot separate these, scoring a random-chance 0.5.
By changing the negative distribution to match the global center bias of the dataset, a model that merely predicts a central focus point will score a 0.5 (random guess). To achieve a high score, the model is forced to accurately predict the unique, image-specific features that deviate from the center (Borji et al. [4], Rahman et al. [7]).
Alternative approaches to this problem include Spatially Binned ROC (spROC), which partitions the image into concentric bins to calculate performance independently across different eccentricities (Wloka et al. [1]), and Farthest-Neighbor AUC (FN-AUC), which prevents deep learning models from gaming metrics via near-binary edge outputs (Jia [8]).
Explicit Spatial Prior Integration
If center bias is an inherent property of how humans view images, models should leverage it—but they must do so explicitly rather than implicitly.
When deep neural networks are trained end-to-end on eye-tracking data, they naturally expend network capacity trying to learn the global 2D Gaussian distribution of human gaze (Pan et al. [9]). This causes two problems: it wastes parameters on a trivial spatial prior, and it leads to high false-positive rates in peripheral regions when deployed in the wild.
To solve this, modern architectures explicitly decouple the spatial prior from feature extraction. Schauerte et al. showed that mathematically integrating an explicit Gaussian center-bias map via multiplicative or linear combinations yields superior results compared to letting a network learn it implicitly [2].
[Input Image] ---> [Deep Feature Extraction] ---\
X ---> [Refined Saliency Map]
[Explicit Spatial Prior] ----/
(Gaussian / Learned Map)
For instance, the Context-Aware Refinement Network (CARNet) utilizes a learned spatial prior that is multiplied directly with the network's deep features to suppress false positives in the periphery (Amirul et al. [10]). Similarly, Linardos et al. integrate an explicit center-bias prior during the final probability calibration phase to ensure continuous, well-calibrated density maps [11]. This can also be enforced during optimization; Liu details how spatial prior regularization can be built directly into the loss function to penalize deviations from known human fixation statistics [12].
By offloading the spatial prior to an explicit mathematical component, the core neural network can focus entirely on what it does best: mapping complex, low-level boundaries and high-level semantic features.
References
- Spatially Binned ROC: A Comprehensive Saliency Metric
- On the Distribution of Salient Objects in Web Images and Its Influence on Salient Object Detection
- Graph-Based Visual Saliency
- Analysis of Scores, Datasets, and Models in Visual Saliency Prediction
- Using Deep Features to Predict Where People Look
- Human Attention in Visual Question Answering: Do Humans and Deep Networks Look at the Same Regions?
- Examining visual saliency prediction in naturalistic scenes
- Revisiting Saliency Metrics: Farthest-Neighbor Area Under Curve
- Shallow and Deep Convolutional Networks for Saliency Prediction
- Salient Object Detection using a Context-Aware Refinement Network
- DeepGaze IIE: Calibrated prediction in and out-of-domain for state-of-the-art saliency modeling
- Deep saliency models : The quest for the loss function
Key Foundational Work
Graph-Based Visual Saliency (2006) — contributed 2 concept elements to this direction.
Interesting objects are visually salient (2008) — contributed 1 concept elements to this direction.