Fusing Residual and Cascade Attention Mechanisms in Voxel-RCNN for 3D Object Detection

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Current 3D object detection methods face challenges in region proposal and small object detection.
• Voxel-based methods often suffer from suboptimal performance due to deep network architectures.

Purpose of the Study:

• To propose an improved 3D object detector, RCAVoxel-RCNN, addressing limitations in existing voxelization techniques.
• To enhance the accuracy and efficiency of 3D object detection, particularly for small-scale objects.

Main Methods:

• Adoption of Voxel-RCNN as a baseline and introduction of RCAVoxel-RCNN.
• Development of a Cascade Attention Network (CAN) for progressive region refinement.
• Implementation of a 3D Residual Network and a Residual Attention Network (RAN) in the Bird's-Eye View (BEV) network.
• Integration of the Squeeze-and-Excitation (SE) attention mechanism for dynamic feature weighting.

Main Results:

• Significant improvements in detection accuracy on the KITTI dataset.
• Achieved 3.34% accuracy increase for cars, 10.75% for pedestrians, and 4.61% for bicycles (KITTI hard level).
• Demonstrated effectiveness of CAN, 3D Residual Network, RAN, and SE attention.

Conclusions:

• The proposed RCAVoxel-RCNN effectively addresses limitations in 3D point cloud voxelization.
• The novel attention and residual network components contribute to superior 3D object detection performance.
• The method shows strong potential for real-world applications requiring precise 3D object recognition.