Reconstructing Superquadrics from Intensity and Color Images.

Area of Science:

• Computer Vision
• 3D Scene Reconstruction
• Deep Learning

Background:

• 3D scene reconstruction is a core computer vision challenge.
• Superquadrics offer a parameter-efficient way to represent complex shapes.
• Prior work successfully used deep learning for superquadric reconstruction from point clouds and depth images.

Purpose of the Study:

• To extend deep learning-based superquadric reconstruction to intensity and color images.
• To develop and evaluate a dedicated convolutional neural network (CNN) for this task.
• To assess the model's performance on various image complexities and compare it to state-of-the-art methods.

Main Methods:

• Utilized a convolutional neural network (CNN) with a ResNet backbone.
• Trained the model to reconstruct single superquadrics from input images (intensity and color).
• Performed qualitative and quantitative analysis, including visualization and error/accuracy metrics.

Main Results:

• The CNN model accurately reconstructs superquadrics from single-object images when one spatial parameter is fixed or inferred.
• Performance degrades only slightly with increased image complexity, such as added textures.
• The proposed model outperforms the current state-of-the-art method.

Conclusions:

• A highly accurate superquadric reconstruction model has been developed.
• The model can reconstruct superquadrics from real-world images of simple objects without additional training.
• This work advances the capabilities of deep learning in 3D shape representation and reconstruction.