An Adaptive Fusion Algorithm for Depth Completion.

Area of Science:

• Computer Vision
• Robotics
• Sensor Fusion

Background:

• LiDAR sensors provide sparse depth data, insufficient for dense perception tasks.
• RGB images offer rich texture but lack direct depth information.
• Existing methods often use simple fusion techniques like concatenation or addition.

Purpose of the Study:

• To develop an adaptive sensor fusion method for LiDAR-based depth completion.
• To improve the accuracy and efficiency of dense depth map generation.
• To overcome limitations of simple feature fusion in multi-modal depth completion.

Main Methods:

• Proposes an adaptive fusion technique using content- and position-aware convolutional kernels.
• Employs an attention network to generate block-specific kernel weights for feature fusion.
• Divides features into blocks and applies attention-generated kernels for multi-modal feature integration.

Main Results:

• Outperformed the state-of-the-art FCFR-Net by 0.01 in inverse Mean Absolute Error (iMAE) on the KITTI dataset.
• Demonstrated superior performance in dense depth completion tasks.
• Achieved a favorable trade-off between computational runtime and accuracy.

Conclusions:

• The proposed adaptive fusion method significantly enhances LiDAR depth completion accuracy.
• The attention-based kernel generation offers a more effective multi-modal feature fusion strategy.
• The method's balance of accuracy and speed makes it suitable for real-time applications.