HCE: Hierarchical Context Embedding for Region-Based Object Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence

Background:

• Current two-stage object detectors use region-wise features that may lack essential context.
• This deficiency hinders the filtering of background noise and recognition of objects with subtle appearances.

Purpose of the Study:

• To introduce a Hierarchical Context Embedding (HCE) framework to improve object detection.
• To enhance the classification capabilities of region-based detectors by mining contextual cues.

Main Methods:

• Developed an image-level categorical embedding module for object-level concept learning.
• Generated novel RoI features using hierarchically embedded context from images and regions.
• Proposed early-and-late fusion strategies (feature and confidence fusion) to leverage these new features.

Main Results:

• The HCE framework demonstrated significant and consistent improvements across various region-based detectors (FPN, Cascade R-CNN, Mask R-CNN, PA-FPN).
• The framework was adapted to one-stage detectors (SSD, RetinaNet, EfficientDet), achieving improved performance.
• HCE proved to be a flexible and generalizable component.

Conclusions:

• The Hierarchical Context Embedding framework effectively integrates contextual information for object detection.
• HCE offers a versatile solution for enhancing both two-stage and one-stage object detection models.