Implicit Negative Sub-Categorization and Sink Diversion for Object Detection.

Area of Science:

• Computer Vision
• Machine Learning

Background:

• Object detection relies on accurate proposal classification.
• The diverse 'background' class poses a challenge in current models.
• Low-probability ground truth scores can hinder detection performance.

Purpose of the Study:

• To improve the proposal classification stage in object detection.
• To enhance the loss function for better performance.
• To address challenges with diverse background classes and low-probability scores.

Main Methods:

• Implicit negative sub-categorization to differentiate background patterns.
• Introduction of a 'sink' class to manage low-value probability scores.
• Utilizing dilated convolution for efficient context extraction.

Main Results:

• Achieved state-of-the-art mean Average Precisions (mAPs) on PASCAL VOC 2007 (84.1%) and 2012 (82.6%).
• Demonstrated a 2.5% improvement on ILSVRC DET compared to ResNet.
• Enhanced performance through strengthened loss functions in the proposal classification stage.

Conclusions:

• The proposed implicit negative sub-categorization and sink diversion effectively improve object detection.
• Dilated convolution aids in extracting valuable context information.
• The refined approach sets a new state-of-the-art in object detection tasks.