Self Paced Deep Learning for Weakly Supervised Object Detection.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Object detection typically requires bounding-box annotations, which are labor-intensive.
• Weakly-supervised object detection uses only image-level labels, posing a challenge for training accuracy.
• Existing iterative methods can introduce false positives due to classifier errors in early stages.

Purpose of the Study:

• To propose a novel training protocol for weakly-supervised object detection using self-paced learning.
• To address the issue of accumulating errors in iterative training frameworks.
• To demonstrate the efficacy of self-paced learning with deep neural networks in an end-to-end pipeline.

Main Methods:

• Implemented a self-paced learning protocol to iteratively train object detectors.
• Focused on selecting reliable image regions and associated bounding boxes for training.
• Adapted the self-paced learning paradigm for deep-network-based classifiers within a Fast-RCNN architecture.

Main Results:

• Achieved state-of-the-art results on benchmark datasets: Pascal VOC 2007, Pascal VOC 2010, and ILSVRC 2013.
• Demonstrated superior performance on ILSVRC 2013 using a low-capacity AlexNet compared to higher-capacity networks.
• Successfully applied self-paced learning to weakly-supervised object detection in an end-to-end deep learning framework.

Conclusions:

• Self-paced learning offers a robust solution for training object detectors with limited annotations.
• The proposed method effectively mitigates false positives common in iterative weakly-supervised approaches.
• This work pioneers the use of self-paced learning in end-to-end deep learning pipelines for object detection.