A Single Stage and Single View 3D Point Cloud Reconstruction Network Based on DetNet.

Area of Science:

• Computer Vision
• 3D Reconstruction
• Deep Learning

Background:

• Inferring 3D object shapes and appearances from single 2D images is a significant challenge.
• Prior work often overlooks detailed 2D image feature extraction and suffers from feature loss during propagation.
• Existing methods may not adequately capture intricate details for accurate 3D reconstruction.

Purpose of the Study:

• To propose a novel single-stage, single-view 3D point cloud reconstruction network, named 3D-SSRecNet.
• To improve the quality of 3D point cloud reconstruction by enhancing 2D feature extraction and minimizing feature loss.
• To achieve more accurate and detailed shape and appearance reconstruction from single images.

Main Methods:

• Developed 3D-SSRecNet, a single-stage network integrating a 2D feature extraction module and a point cloud prediction module.
• Utilized DetNet as the backbone for the 2D feature extraction network to capture finer image details.
• Employed Exponential Linear Unit (ELU) activation and a combined Chamfer Distance (CD) and Earth Mover's Distance (EMD) loss function for point cloud generation.

Main Results:

• 3D-SSRecNet demonstrated superior performance in 3D point cloud reconstruction compared to state-of-the-art methods.
• Experiments on ShapeNet and Pix3D datasets validated the effectiveness of the proposed network.
• The single-stage design effectively reduced feature loss, leading to improved reconstruction quality.

Conclusions:

• The proposed 3D-SSRecNet effectively reconstructs 3D point clouds from single images with enhanced shape and appearance.
• The network's architecture and component choices contribute to overcoming limitations in existing 3D reconstruction techniques.
• 3D-SSRecNet represents a significant advancement in single-view 3D reconstruction.