Laplacian Pyramid Neural Network for Dense Continuous-Value Regression for Complex Scenes.

Area of Science:

• Computer Vision
• Deep Learning
• Image Analysis

Background:

• Dense continuous-value regression (DCR) tasks in computer vision aim to predict pixel-wise continuous values from single images.
• Existing deep convolutional neural network approaches struggle to balance global structure preservation with fine object detail accuracy in complex scenes.

Purpose of the Study:

• To develop a universal and effective neural network architecture for DCR tasks.
• To improve the simultaneous preservation of global scene structure and fine object details.

Main Methods:

• Proposed Laplacian Pyramid Neural Network (LAPNet) utilizing a Laplacian Pyramid Decoder (LPD) for multiscale signal reconstruction.
• Integrated an Adaptive Dense Feature Fusion (ADFF) module for adaptive fusion of multiscale image features.
• Employed a residual refinement module to progressively enhance high-frequency details at each pyramid level.

Main Results:

• Achieved new state-of-the-art performance on monocular depth estimation (NYU-D V2, KITTI).
• Demonstrated superior results in single-image height estimation (Urban Semantic 3D).
• Showcased effectiveness in density map estimation for crowd counting across four benchmarks.

Conclusions:

• LAPNet is a universal and effective architecture for various dense continuous-value regression tasks.
• The proposed LPD and ADFF modules significantly enhance the reconstruction of complex scene signals.
• The method excels in both qualitative and quantitative evaluations, outperforming existing approaches.