Revisiting Siamese-Based 3D Single Object Tracking With a Versatile Transformer.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• 3D Single Object Tracking (SOT) is crucial for applications like autonomous driving and planning.
• Challenges in 3D SOT include sparse point clouds and complex influencing factors.
• Existing methods struggle to effectively process the intricacies of 3D spatial data.

Purpose of the Study:

• To propose a novel Versatile Point Tracking Transformer (VPTT) method for effective 3D SOT.
• To leverage transformer architectures for improved object tracking in sparse 3D point clouds.
• To develop a versatile and efficient end-to-end trainable model for 3D tracking tasks.

Main Methods:

• A siamese-based tracking paradigm utilizing object guidance from template to search point clouds.
• Implementation of self- and cross-attention mechanisms within a deep transformer network (VerFormer).
• Incorporation of four specialized matching operations and intermediate output forecasting (mask probability, distance, heading angle).

Main Results:

• VPTT demonstrates state-of-the-art tracking performance on challenging benchmarks (KITTI, nuScenes, Waymo).
• The method achieves high efficiency, running at approximately 62 FPS on siamese-based tracking.
• The specialized design enables effective end-to-end training for 3D tracking.

Conclusions:

• The proposed VPTT method offers a significant advancement in 3D Single Object Tracking.
• Transformer-based approaches are highly effective for handling sparse point cloud data in tracking.
• VPTT provides a versatile and efficient solution for real-world 3D SOT applications.