Detect2Rank: Combining Object Detectors Using Learning to Rank.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection algorithms have limitations due to specific assumptions on appearance and imaging conditions, lacking universality.
• The proliferation of diverse object detectors necessitates effective methods for their selection and combination.

Purpose of the Study:

• To propose a framework for learning how to combine multiple object detectors.
• To enhance object detection accuracy by exploiting correlations between detectors using contextual features.

Main Methods:

• A framework that integrates single detectors such as Deformable Part Models (DPM), Color Names (CN), and Ensemble of Exemplar-SVMs (EES).
• Utilizes high-level contextual features to exploit correlations among detectors for generating a combined detection list.

Main Results:

• The proposed framework significantly outperforms individual detectors: DPM by 8.4%, CN by 6.8%, and EES by 17.0% on PASCAL VOC07.
• Achieved superior results on PASCAL VOC10: DPM (6.5%), CN (5.5%), and EES (16.2%).
• Demonstrated improved performance (2.4%) over the state-of-the-art RCNN by combining it with other detectors.

Conclusions:

• The proposed framework offers a universal approach to object detection by effectively combining diverse detectors.
• Contextual features play a crucial role in enhancing the performance of combined object detection systems.
• The method shows no constraints on the types of detectors that can be integrated, offering flexibility.