[Paper Review] One-Bit Object Detection: On learning to localize objects with minimal supervision.

Hyun Oh Song, Ross Girshick+4 more|arXiv (Cornell University)|Mar 5, 2014

Advanced Neural Network Applications12 citations

TL;DR

This paper proposes a one-bit object detection method that localizes objects using only image-level labels indicating presence or absence. By combining a discriminative submodular cover for positive window discovery with a smoothed latent SVM and Quasi-Newton optimization, the approach achieves a ~70% relative improvement in average precision over state-of-the-art methods on standard benchmarks.

ABSTRACT

Learning to localize objects with minimal supervision is an important problem in computer vision, since large fully annotated datasets are extremely costly to obtain. In this paper, we propose a new method that achieves this goal with only image-level labels of whether the objects are present or not. Our approach combines a discriminative submodular cover problem for automatically discovering a set of positive object windows with a smoothed latent SVM formulation. The latter allows us to leverage efficient Quasi-Newton optimization techniques. Our experiments demonstrate that the proposed approach provides approximately 70% relative improvement in average precision over the current state of the art on standard benchmark datasets.

Motivation & Objective

• To address the high cost of fully annotated object detection datasets by enabling object localization with only image-level labels.
• To develop a weakly supervised learning method that discovers relevant object windows without bounding box annotations.
• To improve localization accuracy under minimal supervision by leveraging submodular optimization and efficient optimization techniques.
• To outperform existing state-of-the-art methods in average precision on standard benchmarks using only presence/absence labels.

Proposed method

• The method uses a discriminative submodular cover to automatically identify a set of positive object windows from image-level labels.
• It formulates the detection problem as a smoothed latent SVM to enable efficient Quasi-Newton optimization.
• The smoothed latent SVM allows end-to-end training while maintaining differentiability and convergence guarantees.
• The approach jointly optimizes for object window selection and detector parameters using weak supervision.
• Submodular optimization ensures diverse and high-quality positive window candidates are selected.

Experimental results

Research questions

• RQ1 Can object localization be effectively learned using only image-level labels indicating object presence or absence?
• RQ2 How can submodular optimization be leveraged to discover meaningful positive object windows under weak supervision?
• RQ3 Can a smoothed latent SVM formulation enable efficient and accurate training under minimal supervision?
• RQ4 To what extent does this method improve average precision compared to existing weakly supervised object detection approaches?

Key findings

• The proposed method achieves approximately 70% relative improvement in average precision over the current state of the art on standard benchmark datasets.
• The use of a discriminative submodular cover significantly enhances the quality of positive window selection from image-level labels.
• The smoothed latent SVM formulation enables effective optimization via Quasi-Newton methods, improving convergence and performance.
• The method demonstrates strong generalization with minimal supervision, outperforming prior weakly supervised detection approaches.