Guided Depth Completion with Instance Segmentation Fusion in Autonomous Driving Applications.

Area of Science:

• Computer Vision
• Robotics
• 3D Scene Understanding

Background:

• Pixel-level depth data is vital for applications like autonomous driving and robotics.
• Current depth completion methods often use resource-intensive networks without object-specific information.
• Sparse depth data from sensors like LiDAR requires sophisticated completion techniques.

Purpose of the Study:

• To develop a more efficient and accurate depth completion framework by incorporating object-level information.
• To leverage image instance segmentation for pixel-level object detection to guide depth prediction.
• To improve the performance of depth completion models by fusing multi-modal data.

Main Methods:

• A two-branch encoder-decoder deep neural network architecture was employed.
• Image instance segmentation was used to detect and localize objects of interest.
• The framework fused object information (type, location), LiDAR, and RGB camera data.
• The method was evaluated on the KITTI dataset.

Main Results:

• The proposed method demonstrated faster training convergence compared to baseline models.
• Improved prediction accuracy was achieved across all evaluation metrics.
• The object-aware approach led to more efficient and effective depth map generation.

Conclusions:

• Integrating object instance segmentation significantly enhances depth completion performance.
• The proposed framework offers a more efficient and accurate solution for dense depth map prediction.
• This object-centric approach holds promise for real-time applications in autonomous systems.