Distance Transform Pooling Neural Network for LiDAR Depth Completion.

Area of Science:

• Computer Vision
• Robotics
• Deep Learning

Background:

• Recovering dense depth maps from sparse sensors like LiDAR is crucial for computer vision and robotics.
• Input sparsity from LiDAR data presents a significant challenge for existing methods.
• Current approaches often require deep, computationally intensive networks to compensate for sparse inputs.

Purpose of the Study:

• To propose a novel module for efficiently processing sparse LiDAR data for depth map recovery.
• To address the key challenge of input sparsity in depth completion tasks.
• To achieve state-of-the-art performance using a lightweight neural network architecture.

Main Methods:

• Introduced a recurrent distance transform pooling (DTP) module to aggregate multi-level information from sparse depth data.
• Developed an error correction (EC) module to handle erroneous LiDAR sensor readings.
• Utilized a simplified ResNet-18 backbone for depth completion, processing multi-level semi-dense depth maps generated by the DTP module.

Main Results:

• The proposed DTP module effectively alleviates the sparsity challenge, enabling a lightweight network to achieve state-of-the-art results.
• Achieved top performance on the KITTI depth completion benchmark using LiDAR data exclusively.
• The EC module successfully prevented the propagation of incorrect sensor values.

Conclusions:

• The recurrent DTP module offers an effective solution for dense depth map recovery from sparse LiDAR data.
• The method demonstrates the potential for efficient and accurate depth completion with lightweight models.
• Future work includes exploring sensor fusion and indoor applications.