研究目的
To recover the 3D shape of texture-less, deformable surfaces from a single image using a data-driven approach, overcoming the limitations of traditional Shape-from-Shading methods.
研究成果
The study demonstrates that a data-driven approach can effectively reconstruct the 3D shape of texture-less, deformable surfaces from a single image, outperforming traditional Shape-from-Shading methods in terms of accuracy and speed. The use of depth- and normal-based representations was found to be more effective than mesh-based ones for this task.
研究不足
The method's performance may degrade when training on one object and testing on a significantly different one, indicating a need for larger and more diverse training datasets. The process of obtaining ground-truth mesh coordinates is more tedious and error-prone than for depth or normal maps.
1：Experimental Design and Method Selection:
The study employs a data-driven approach using a deep autoencoding architecture (SegNet) to regress from the image to various 3D representations (meshes, normals, depth maps).
2：Sample Selection and Data Sources:
A new large dataset of real deforming surfaces (cloth, T-shirt, sweater, hoody, paper) under varying lighting conditions was created, comprising 26500 image-normals-depth triplets.
3：List of Experimental Equipment and Materials:
Microsoft Kinect camera for capturing synchronized RGB images and depth maps, incandescent lamps for lighting.
4：Experimental Procedures and Operational Workflow:
The process involves capturing sequences of deformable surfaces, preprocessing the data (segmentation, white balancing, cropping), and training the network to predict 3D shapes from single images.
5：Data Analysis Methods:
Evaluation metrics include mean vertex-wise Euclidean distance for mesh coordinates, alignment of depth maps using Procrustes transformation, and angular errors for normal maps.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
