Bi-Att3DDet: Attention-Based Bi-Directional Fusion for Multi-Modal 3D Object Detection

Area of Science:

• Computer Vision
• Autonomous Driving Systems
• Sensor Fusion

Background:

• Multi-modal 3D object detection is crucial for autonomous driving.
• Existing methods struggle with effective fusion of LiDAR and RGB image features.
• Inadequate capture of structural information in Region of Interest (RoI) features limits performance.

Purpose of the Study:

• To propose Bi-Att3DDet, a multi-modal sensor fusion network for improved 3D object detection.
• To enhance the utilization of complementary information between depth and semantic texture features.
• To better capture structural information within RoI features for more accurate detections.

Main Methods:

• Developed a Self-Attentive RoI Feature Extraction module (SARoIFE) using self-attention mechanisms.
• Implemented a Feature Bidirectional Interactive Fusion module (FBIF) for LiDAR and pseudo RoI features.
• Employed comprehensive experiments on the KITTI dataset to validate the proposed method.

Main Results:

• Achieved a 1.55% improvement on the hard difficulty level.
• Secured a 0.19% improvement in mean Average Precision (mAP) on the test dataset.
• Demonstrated effective fusion of complementary sensor information and improved RoI feature extraction.

Conclusions:

• Bi-Att3DDet significantly enhances multi-modal 3D object detection performance.
• The proposed SARoIFE and FBIF modules effectively address limitations in feature fusion and structural information capture.
• The method shows strong potential for real-world autonomous driving applications.