Prior scene context reshapes feature reliance during rapid perception

By combining eye tracking with feature-based encoding models, this study demonstrates that prior scene context facilitates rapid face detection by shifting the perceptual strategy from reliance on sensory-driven features to expectation-based spatial guidance, a change evident even in the first eye movement.

The Gist

Imagine your brain is a detective trying to spot a specific person (a face) hidden inside a crowded, chaotic room (a natural scene). The paper asks: Does knowing what the room looks like before you enter help the detective find the person faster?

To answer this, the researchers set up two "games" with 38 volunteers each. They used special eye-tracking glasses to see exactly where people looked and how long it took them to spot a face.

The Two Games

1. Game 1 (The Preview): Before the real picture appeared, participants saw a "faceless" version of the same scene. It was like looking at a map of the room before walking in.
2. Game 2 (The Tunnel): This time, they saw the faceless map and a moving window that only let them see a small part of the real scene at a time. This made the job much harder, forcing them to rely even more on that initial map.

What They Found

The results showed that having that "faceless map" (the prior context) made people find faces much faster, especially when the faces were hard to spot.

Here is the most interesting part: The brain used the map immediately. Even the very first time the person's eyes moved to look at the scene, they were already guided by what they expected to see. They didn't just scan randomly; they went straight to where a face was likely to be based on the preview.

The Detective's Toolkit: Two Ways to Find a Clue

To understand how the brain did this, the researchers built a computer model with two types of "clues":

• Clue Type A (The Sensory Input): This is what the eyes actually see right now—the shape of a nose, the color of skin, the contrast of features. It's like looking at a blurry photo and trying to guess what it is based purely on the pixels.
• Clue Type B (The Expectation): This is the "spatial prior." It's the brain's guess of where a face should be based on the scene's layout. It's like knowing, "In a kitchen, people usually stand near the counter, not floating in the middle of the ceiling."

The Big Shift

The study found that both types of clues helped the brain work, but the balance changed depending on whether the preview was available.

• Without the preview: The brain relied heavily on Clue Type A. It had to squint and analyze every pixel of sensory data to find the face.
• With the preview: The brain leaned much more heavily on Clue Type B. It trusted its expectation of where the face would be. Because it knew where to look, it didn't need to work as hard to process the raw sensory details.

The Bottom Line

Think of it like searching for your keys. If you have no idea where you left them, you have to frantically look at every single object on the table (relying on sensory input). But if you remember, "I usually put my keys on the hook by the door," you walk straight to the hook and grab them without even looking at the rest of the table (relying on expectation).

This paper proves that when we have a "map" of the scene beforehand, our brains instantly switch from a "search mode" to a "targeted mode," using our expectations to guide our eyes before we even fully process what we see.

Technical Summary

Technical Summary: Prior Scene Context Reshapes Feature Reliance during Rapid Perception

Problem Statement
Human visual perception is known to be a synthesis of immediate sensory input and prior knowledge or expectations. However, the specific mechanisms by which prior contextual information influences rapid visual processing remain unclear. This study addresses the question of how prior contextual information modulates the reliance on sensory features versus spatial expectations during the rapid detection of faces in natural scenes.

Methodology
The research employed a combination of eye tracking and feature-based encoding models across two experiments involving 38 participants each. The core task was rapid face detection within natural scenes.

• Experiment 1: Manipulated the presence of "faceless" scene previews to establish a prior context before the target image appeared.
• Experiment 2: Replicated the conditions of the first experiment but added a moving-window paradigm to restrict peripheral visual input, thereby forcing a greater reliance on prior information.

To quantify the information guiding behavior, the authors modeled detection latencies using a set of image-based predictors categorized into two classes:

1. Sensory Information: Capturing low-level and high-level visual features.
2. Scene-Derived Spatial Prior: Specifically, the expected location of a face based on the scene context.

Key Results

• Facilitation by Context: Across both experiments, the availability of prior scene context facilitated face detection, with the effect being most pronounced for challenging images.
• Temporal Dynamics: This facilitation was evident in the very first eye movement, indicating that scene previews shape perceptual strategies from the onset of processing rather than through later cognitive adjustments.
• Predictive Modeling: Both sensory predictors and spatial priors successfully explained variance in detection latencies across images.
  • Among sensory predictors, the difference in deep neural network (DNN) responses induced by the presence of a face provided the strongest out-of-sample prediction of detection latency.
  • Shift in Reliance: Critically, when scene previews were available, the contribution of the spatial prior to the model increased, while the reliance on sensory-driven features generally decreased.

Significance and Claims
The paper claims that prior scene context fundamentally alters the computational balance used for rapid face detection. Specifically, the findings indicate that context shifts the information processing strategy from a dominance of sensory-driven features toward expectation-based spatial guidance. This suggests that the visual system does not merely add prior knowledge to sensory data but actively reweights the reliance on different information sources depending on the availability of contextual cues.