Cost-Effective Object Detection: Active Sample Mining With Switchable Selection Criteria.

Area of Science:

• Computer Science
• Machine Learning
• Artificial Intelligence

Background:

• Leveraging large-scale unlabeled or partially labeled data is crucial for training learning systems.
• Existing active learning (AL) methods often overlook majority samples, focusing only on uncertain minority samples.

Purpose of the Study:

• To introduce a principled active sample mining (ASM) framework for cost-effective training of object detectors.
• To demonstrate that mining unlabeled majority data is key to powerful object detection with minimal user effort.

Main Methods:

• Developed an ASM framework with a switchable sample selection mechanism for manual annotation (AL) or automatic pseudolabeling.
• Implemented a novel self-learning process compatible with mini-batch training for object detection.
• Utilized deep neural networks to estimate labels and confidences for unlabeled samples, assigning temporary pseudolabels for fine-tuning.

Main Results:

• The ASM framework achieves performance comparable to alternative methods using significantly fewer annotations.
• Demonstrated effectiveness on PASCAL VOC 2007/2012 benchmarks.
• The method is suitable for object categories unseen during the learning process.

Conclusions:

• Active sample mining is a cost-effective strategy for training powerful object detectors.
• The proposed ASM framework effectively balances manual annotation and self-learning for improved efficiency.
• This approach enhances object detection capabilities, especially with limited labeled data.