Top-down control of visual attention in object detection

Oliva, Aude; Torralba, Antonio; Castelhano, Monica S.; Henderson, John M. · 2004 · OpenAlex-citations

DOI: 10.1109/icip.2003.1246946

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the limitation of current computational models of visual attention, which primarily rely on bottom-up information (local image features) and ignore scene context. The authors argue that humans utilize top-down contextual information to facilitate object detection by directing attention to diagnostic regions. The research aims to validate a computational model that integrates global scene configuration with local saliency to predict where observers look when searching for specific objects, specifically people, in natural scenes. The study proposes a probabilistic model of attention guidance that combines local saliency with contextual priors. Local saliency is defined as the inverse likelihood of finding specific low-level features (derived from steerable pyramids) within the image, approximated using a Gaussian distribution. Contextual priors are computed by learning the relationship between global scene features and target locations from a database of images. Contextual features are extracted by reducing the dimensionality of local features via absolute value transformation, subsampling, and Principal Component Analysis (PCA), resulting in a 60-dimensional vector. The model calculates a contextual saliency map, $S_c(x)$, which modulates local saliency based on the probability of finding the target at a given location given the context, without requiring specific knowledge of the target’s appearance. To evaluate the model, the authors compared its predicted fixation patterns against human eye movement data. Eight subjects performed a visual search task, counting people in 36 real-world scenes while their eye movements were recorded. The analysis focused on the first seven fixations. The predicted fixation patterns from three conditions were compared to human data: a purely random pattern, a standard bottom-up saliency model (Itti et al.), and the proposed contextual model ($S_c$). The similarity between fixation patterns was measured using squared differences, normalized against the average distance between human subjects. The results demonstrated that the contextual model’s fixation patterns most closely resembled those of human observers. Pure bottom-up saliency models performed worse than the contextual model but were still more similar to human data than random fixations. There was no statistical difference in performance between the standard Itti et al. model and the probabilistic definition of local saliency used in this study. These findings validate the proposition that top-down information from visual context modulates the saliency of image regions during object detection. The authors conclude that contextual information provides an efficient, low-cost shortcut for object detection systems by pre-selecting relevant regions based on scene statistics, thereby reducing the computational expense of exhaustive search procedures.

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-25; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• visual search