研究目的
To present a new projection mapping framework based on BRDF reconstruction for more realistic projection results by enhancing Augmented Reality effects, addressing the lack of attention to material properties in previous methods.
研究成果
The proposed framework effectively reconstructs BRDF and geometric shape from a single RGBD image using CNNs, reducing color deviation and improving realism in projection mapping compared to hand-crafted methods. Future work should focus on real-time performance, more complex rendering models, and handling dynamic objects.
研究不足
The framework is not real-time due to complex computations; the BRDF model used (Blinn-Phong) is not highly complex, limiting realism; relies on a fixed and known light source; depth information from Kinect may have errors.
1：Experimental Design and Method Selection:
The framework uses a monocular vision approach with a single RGBD image to reconstruct BRDF parameters and geometric shape using two CNNs (PR-Net for reflectance map and PN-Net for normal map), inspired by prior work but adapted for single-image input.
2：Sample Selection and Data Sources:
A custom dataset of 600 RGBD images under fixed projector illumination, featuring various objects with at least three materials per scene (e.g., skin, plaster, clothes), split into 480 training and 120 testing images.
3：List of Experimental Equipment and Materials:
Hardware includes a projector (Panasonic PT-BX40), a Kinect
4：0 depth camera, and a desktop PC (Intel i7 processor, 16GB RAM, Nvidia GeForce GTX 1080 GPU). Software involves OpenGL Shader for rendering and custom CNN implementations. Experimental Procedures and Operational Workflow:
Calibrate Kinect and projector using a chessboard pattern; capture RGBD image; preprocess data; predict reflectance and normal maps using CNNs with prior normal map from depth; render content with Blinn-Phong model; project onto object.
5：Data Analysis Methods:
Evaluate accuracy using absolute (acc0) and relative (acc5) pixel difference metrics; compare with hand-crafted Lambertian materials; use loss functions combining mean absolute error and mean square error.
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
