[Paper Review] Unsupervised deep object discovery for instance recognition

Oriane Siméoni, Ahmet İşcen+3 more|arXiv (Cornell University)|Sep 14, 2017

Advanced Image and Video Retrieval Techniques39 references3 citations

TL;DR

This paper proposes an unsupervised deep object discovery method that identifies salient regions in images using a centrality measure on a nearest neighbor graph built from regional CNN features. By focusing on discriminative and common patterns, the method improves instance-level retrieval, especially for small objects in cluttered scenes, outperforming global descriptors in large-scale image retrieval under severe background clutter.

ABSTRACT

Severe background clutter is challenging in many computer vision tasks, including large-scale image retrieval. Global descriptors, that are popular due to their memory and search efficiency, are especially prone to corruption by such a clutter. Eliminating the impact of the clutter on the image descriptor increases the chance of retrieving relevant images and prevents topic drift due to actually retrieving the clutter in the case of query expansion. In this work, we propose a novel salient region detection method. It captures, in an unsupervised manner, patterns that are both discriminative and common in the dataset. Saliency is based on a centrality measure of a nearest neighbor graph constructed from regional CNN representations of dataset images. The descriptors derived from the salient regions improve particular object retrieval, most noticeably in a large collections containing small objects.

Motivation & Objective

• To address the challenge of severe background clutter in large-scale image retrieval, which degrades performance of global descriptors.
• To develop an unsupervised method that identifies salient regions that are both discriminative and common across a dataset.
• To improve instance recognition by deriving descriptors from these salient regions instead of global image features.
• To enhance retrieval performance, particularly for small objects, in datasets with high background clutter.
• To reduce topic drift during query expansion by minimizing the influence of irrelevant background content.

Proposed method

• Construct a nearest neighbor graph using regional CNN feature representations extracted from images in the dataset.
• Compute a centrality measure—specifically, a variant of PageRank—on the nearest neighbor graph to identify salient regions.
• Use the centrality scores to localize regions that are both common and discriminative across the dataset.
• Extract deep features from the salient regions to form improved image descriptors for retrieval.
• Train and evaluate the method in an unsupervised manner without requiring bounding box annotations or category labels.
• Apply the saliency-based descriptors to instance-level image retrieval, comparing performance against global descriptor baselines.

Experimental results

Research questions

• RQ1 Can unsupervised saliency detection based on graph centrality improve instance retrieval in cluttered datasets?
• RQ2 How effective is the proposed method in identifying discriminative and common regions without supervision?
• RQ3 Does descriptor learning from salient regions outperform global descriptors in large-scale image retrieval with small objects?
• RQ4 To what extent does the method reduce topic drift during query expansion in cluttered environments?
• RQ5 How does the performance vary across datasets with varying levels of background clutter?

Key findings

• The proposed method significantly improves instance-level retrieval performance, especially in datasets with high background clutter.
• Saliency detection based on graph centrality successfully identifies regions that are both discriminative and common across the dataset.
• Descriptors derived from salient regions outperform global descriptors in large-scale retrieval, particularly for small objects.
• The method reduces topic drift during query expansion by minimizing the retrieval of irrelevant background content.
• The unsupervised nature of the approach enables effective salient region discovery without any annotated data.
• The method demonstrates consistent gains across multiple benchmark datasets, confirming its robustness to clutter and object scale.