Monocular Depth Estimation Using a Laplacian Image Pyramid with Local Planar Guidance Layers

Area of Science:

• Computer Vision
• Deep Learning
• 3D Reconstruction

Background:

• Accurate depth estimation from 2D images is crucial but challenging, especially at object boundaries.
• Existing deep learning methods for monocular depth estimation struggle with precise boundary prediction.

Purpose of the Study:

• To develop a depth estimation network that emphasizes precise object boundaries for improved depth map accuracy.
• To enhance the sophistication of depth prediction by focusing on fine details.

Main Methods:

• Proposed a novel encoder-decoder depth estimation network utilizing Laplacian pyramid and local planar guidance.
• Employed ConvNeXt networks as the backbone, constructing the deep neural network (DNN) using only convolutions.
• Integrated Laplacian pyramid and local planar guidance during the upsampling phase to refine object boundaries.

Main Results:

• Achieved an absolute relative error (Abs_rel) of 0.054 and root mean square error (RMSE) of 2.252 on the KITTI dataset.
• Attained an Abs_rel of 0.102 and RMSE of 0.355 on the NYU Depth V2 dataset.
• Demonstrated state-of-the-art performance, ranking fifth on the KITTI Eigen split and eighth on the NYU Depth V2 benchmark.

Conclusions:

• The proposed network effectively improves depth estimation accuracy by focusing on object boundaries.
• The integration of Laplacian pyramid and local planar guidance enhances the detail and precision of predicted depth maps.
• The method represents a significant advancement in monocular depth estimation, achieving competitive results on established datasets.