Deformable Part Region Learning and Feature Aggregation Tree Representation for Object Detection.

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) have advanced object detection.
• Object detectors struggle with accuracy due to low feature discriminability from geometric object variations.
• Existing methods often require extensive labeled data for part models.

Purpose of the Study:

• To propose Deformable Part Region (DPR) learning for adaptable object part detection.
• To introduce a Feature Aggregation Tree (FAT) for learning more discriminative Region of Interest (RoI) features.
• To develop a novel cascade architecture for iterative refinement of detection tasks.

Main Methods:

• Deformable Part Region (DPR) learning allows decomposed object parts to deform with geometric transformations.
• Part model losses are designed for detection and segmentation, enabling unsupervised learning of geometric parameters.
• A Feature Aggregation Tree (FAT) with spatial and channel attention aggregates part RoI features for enhanced semantic representation.
• A cascade architecture iteratively refines object detection and segmentation.

Main Results:

• The proposed DPR and FAT networks, integrated into a cascade architecture, significantly improve object detection and segmentation.
• The Cascade D-PRD model achieved 57.9 box AP on MSCOCO and PASCAL VOC datasets using a Swin-L backbone.
• Extensive ablation studies confirmed the effectiveness and utility of the proposed methods for large-scale object detection.

Conclusions:

• The DPR learning approach effectively handles object geometric variations without requiring explicit part ground truth.
• The FAT network enhances feature discriminability by aggregating part-level information through a bottom-up tree structure.
• The developed cascade architecture offers a robust framework for state-of-the-art object detection and segmentation performance.