Refined Voting and Scene Feature Fusion for 3D Object Detection in Point Clouds

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• 3D object detection in lidar point clouds is crucial for visual world understanding.
• Existing Hough voting methods struggle with inaccurate centroid prediction, leading to redundant bounding boxes.

Purpose of the Study:

• To propose a novel network, RSFF-Net, for accurate indoor 3D object detection.
• To enhance 3D object detection by refining voting mechanisms and integrating scene context.

Main Methods:

• RSFF-Net utilizes a geometric function module to capture object features.
• A refined voting module fuses geometric features with coarse votes for improved centroid prediction.
• A scene constraint module incorporates object-scene associations to guide detection.

Main Results:

• RSFF-Net demonstrates competitive performance on indoor 3D object detection benchmarks.
• The refined voting and scene feature fusion effectively reduce redundant bounding box generation.

Conclusions:

• RSFF-Net offers a simple yet effective approach to indoor 3D object detection.
• Integrating geometric features and scene context significantly improves detection accuracy and reduces errors.