[Paper Review] Collaborative Receptive Field Learning

Shu Kong, Zhuolin Jiang+1 more|arXiv (Cornell University)|Feb 2, 2014

Visual and Cognitive Learning Processes31 references3 citations

TL;DR

This paper proposes Collaborative Receptive Field Learning (coRFL), a weakly supervised method that discovers category-specific, foreground-focused image regions (receptive fields) across multiple images using a submodular optimization framework. By introducing Pyramid-Error Distance (PED) for measuring region similarity and a nonparametric classifier, coRFL improves object categorization performance without requiring bounding box annotations, effectively purifying training data by retaining only meaningful object regions.

ABSTRACT

The challenge of object categorization in images is largely due to arbitrary translations and scales of the foreground objects. To attack this difficulty, we propose a new approach called collaborative receptive field learning to extract specific receptive fields (RF's) or regions from multiple images, and the selected RF's are supposed to focus on the foreground objects of a common category. To this end, we solve the problem by maximizing a submodular function over a similarity graph constructed by a pool of RF candidates. However, measuring pairwise distance of RF's for building the similarity graph is a nontrivial problem. Hence, we introduce a similarity metric called pyramid-error distance (PED) to measure their pairwise distances through summing up pyramid-like matching errors over a set of low-level features. Besides, in consistent with the proposed PED, we construct a simple nonparametric classifier for classification. Experimental results show that our method effectively discovers the foreground objects in images, and improves classification performance.

Motivation & Objective

• To address the challenge of arbitrary object translations and scales in image categorization by learning category-specific receptive fields without precise object localization.
• To develop a collaborative framework that jointly identifies salient image regions across multiple images of the same category using only category-level labels.
• To improve image classification by purifying training data through retention of only the most discriminative, foreground-focused receptive fields.
• To design a robust, nonparametric similarity metric (PED) for comparing image regions with varying sizes and spatial configurations.

Proposed method

• The method models receptive field discovery as a submodular function maximization problem over a similarity graph constructed from candidate receptive fields.
• It introduces Pyramid-Error Distance (PED), a nonparametric metric that computes region similarity by summing pyramid-like matching errors across multiple levels of low-level SIFT features.
• A Gaussian kernel with a learned scale parameter σ is applied to transform the PED-based dissimilarity graph into a similarity graph for graph construction.
• The submodular objective function incorporates constraints to ensure each image contributes at least one positive receptive field (via λ₁) and a center-bias to favor central object locations (via λ₂).
• A greedy algorithm is used to solve the submodular optimization, providing performance-guaranteed solutions for selecting the most representative receptive fields.
• A nonparametric classifier is designed based on PED to match query images with the purified training set of selected receptive fields.

Experimental results

Research questions

• RQ1 Can we effectively discover foreground object regions across multiple images of the same category using only weak category-level labels?
• RQ2 How can we measure the similarity between image regions of varying sizes and spatial configurations in a way that supports robust region matching?
• RQ3 To what extent can submodular optimization with vision-based priors improve the selection of discriminative receptive fields for object categorization?
• RQ4 Can a nonparametric classifier based on the proposed PED metric achieve competitive performance with minimal assumptions on data distribution?

Key findings

• The proposed coRFL framework successfully identifies foreground object regions across multiple images without requiring bounding box annotations, significantly improving training set quality.
• The Pyramid-Error Distance (PED) metric effectively captures regional similarity across diverse scales and translations, outperforming standard feature-based distance measures in region matching.
• Experiments on Caltech101 and synthetic data show that coRFL improves classification accuracy by focusing on salient, category-specific image regions.
• The method demonstrates robustness to parameter variations, particularly with τ and λ₂, and achieves optimal performance with σ = 0.3 in the Gaussian kernel for similarity graph construction.
• The submodular optimization framework enables efficient, performance-guaranteed selection of receptive fields, making it scalable for large-scale image datasets.
• The nonparametric classifier based on PED achieves strong classification results, validating the effectiveness of the purified training set.