WCNN3D: Wavelet Convolutional Neural Network-Based 3D Object Detection for Autonomous Driving.

Area of Science:

• Computer Vision
• Deep Learning
• Signal Processing

Background:

• Standard Convolutional Neural Networks (CNNs) suffer information loss due to pooling operations, hindering 3D object detection in autonomous driving.
• Existing methods often struggle with comprehensive environmental perception due to limitations in feature representation.

Purpose of the Study:

• To design a novel 3D object detection network that avoids information loss by excluding pooling operations.
• To improve feature representation and receptive field size using wavelet transforms and multi-frequency coefficients.
• To develop a lightweight yet high-performance model for autonomous driving perception.

Main Methods:

• A wavelet-multiresolution-analysis-based network incorporating discrete wavelet transform (DWT) and inverse wavelet transform (IWT).
• Utilized lower and higher frequency coefficients as filters instead of single filters in standard convolution.
• Implemented skip connections with element-wise summation for efficient feature reuse and reduced computational load.

Main Results:

• The proposed model successfully avoids information loss associated with pooling operations.
• Wavelet-based filters demonstrated enhanced feature capture and an enlarged receptive field compared to standard convolutions.
• Experimental results on the KITTI benchmark showed performance improvements of up to 14% over PointPillars with fewer trainable parameters.

Conclusions:

• The wavelet-based 3D object detection network offers a promising alternative to standard CNNs for autonomous driving.
• The model achieves a balance between performance and computational efficiency, enabling lightweight yet effective perception systems.
• Further research can explore different wavelet families and decomposition levels for optimized performance.