Large-Scale Object Detection in the Wild With Imbalanced Data Distribution, and Multi-Labels.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning performance heavily relies on training data volume.
• Large-scale datasets like Open Images offer opportunities but present challenges.
• The Open Images dataset suffers from label noise and class imbalance due to its collection method.

Purpose of the Study:

• To quantitatively analyze label-related problems in large-scale object detection datasets.
• To propose and demonstrate effective solutions for multi-label and imbalanced data issues.
• To advance the state-of-the-art performance on the Open Images object detection benchmark.

Main Methods:

• Developed a concurrent softmax approach to manage multi-label object detection.
• Implemented a soft-balance sampling method combined with a hybrid training scheduler to address label imbalance.
• Designed an ensemble mechanism to further enhance single-model performance.

Main Results:

• Achieved a single-model mean average precision (mAP) of 60.90% on the Open Images test set, a 3.34 point improvement.
• The ensemble model reached an overall mAP of 67.17%, surpassing previous benchmarks by 4.29 points.
• Demonstrated significant performance gains by effectively tackling label multiplicity and distribution imbalance.

Conclusions:

• The proposed concurrent softmax and soft-balance sampling effectively resolve major challenges in large-scale object detection.
• Ensemble methods provide substantial performance boosts, setting new state-of-the-art results on the Open Images dataset.
• This work offers a comprehensive solution for improving deep learning models trained on complex, large-scale datasets.