Enhanced 3D shape recovery using the neural-based hybrid reflectance model

Area of Science:

• Computer Vision
• Computer Graphics
• Artificial Intelligence

Background:

• Real-world surfaces exhibit hybrid reflectance properties, combining Lambertian and specular characteristics.
• Existing hybrid reflectance models struggle with noise, strong specularities, and unknown reflectivity.
• Accurate surface reflectance modeling is crucial for 3D shape recovery techniques like shape from shading.

Purpose of the Study:

• To propose a novel neural-based hybrid reflectance model that addresses limitations of existing methods.
• To optimize the reflectance model by learning weights and parameters of feedforward neural networks and radial basis function networks.
• To enhance 3D object shape recovery using the shape from shading technique with the proposed model.

Main Methods:

• Developed a hybrid reflectance model integrating feedforward neural networks (FNNs) and radial basis function networks (RBFNs).
• Employed a learning-based approach to optimize the weights and parameters of the hybrid FNN-RBFN structure.
• Applied the optimized reflectance model within a shape from shading framework for 3D shape recovery.

Main Results:

• The proposed neural-based hybrid reflectance model demonstrated improved performance in handling noise and specular effects.
• Experimental results with synthetic and real images validated the model's effectiveness across various conditions.
• Successful 3D object shape recovery was achieved using the optimized reflectance model.

Conclusions:

• The novel neural-based hybrid reflectance model offers a robust solution for complex surface reflectance.
• The method enhances the accuracy and reliability of shape from shading techniques.
• This approach advances the field of 3D reconstruction by addressing key challenges in surface modeling.