Improving the Speed of Virtual Rear Projection: A GPU-Centric Architecture Matthew Flagg, Jay Summet, James M. Rehg GVU Center College of Computing Georgia Institute of Technology
Matthew Flagg © Ubiquitous Interactive Displays Every flat surface can be an interactive display
Matthew Flagg © VRP: Shadow Elimination Single Projector Case
Matthew Flagg © Half power shadows Shadow Elimination Double Projector Case Passive VRP
Matthew Flagg © Shadow Elimination Boosting projector outputs Proportional feedback law
Matthew Flagg © Occluder Light Suppression Detecting occluded pixels
Matthew Flagg © Detecting occluded pixels Occluder Light Suppression
Matthew Flagg © Detecting occluded pixels Occluder Light Suppression
Matthew Flagg © Detecting occluded pixels Occluder Light Suppression Nonlinear feedback law
Matthew Flagg © Virtual Rear Projection Show ICCV’03 demo video
Matthew Flagg © Challenges for VRP High image quality Seams between display regions projected by different projectors Photometric Uniformity Fast Compensation Avoid perception of shadows caused by system lag Image processing required to ensure high image quality
Matthew Flagg © Camera view of screen must be unobstructed Requires reference image capture before occlusion Cannot be co-located with projector Shadows still perceptible Shadow detection image processing performed on CPU Limitations With Previous Work
Matthew Flagg © Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 New Approach
Matthew Flagg © New Approach Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 IR backlit camera image
Matthew Flagg © Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 New Approach Turn off occluded pixels
Matthew Flagg © New Approach Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 Occluder Light Suppression
Matthew Flagg © Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Shadow Eliminator Blinding Light Suppressor Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 New Approach Turn on occluded pixels with second projector
Matthew Flagg © New Approach Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 Shadow Elimination
Matthew Flagg © New Approach Detect Occluders, Not Shadows Co-locate projector with camera Active IR imaging Based on work by Tan and Pausch CHI’02 Projector Roles: Blinding Light Suppressor Shadow Eliminator Image Processing on GPU Pixel Shaders Render-To-Texture with DirectX9.0 Shadow Elimination and Occluder Light Suppression
Matthew Flagg © Fast Compensation: GPU-Centric Approach
Matthew Flagg © Fast Compensation: GPU-Centric Approach 1. Warping, background subtraction 2. Median filtering and dilation for inter-frame tolerance 3. Gaussian blur for blending 4. Compositing and warping
Matthew Flagg © Pixel Shader Pipeline camera texture background texture render texture 1 render texture 2 back buffer display image (A)(B)(C)
Matthew Flagg © Addressing Image Quality LAM for left projector LAM for right projector Luminance Attenuation Maps (LAMs) Simple feedback-based approach to accommodate non-linearities of projector- camera Seam Blending seam – no blending seam – with blending
Matthew Flagg © Virtual Rear Projection Results Play Video
Matthew Flagg © Virtual Rear Projection Results System ComponentLatency Camera Capture to PC-Memory9.09ms PC-Memory to GPU-Memory1.70ms Pixel Shaders2.14ms Projector40.27ms Total Latency53.20ms Image processing speed increased from 15Hz to 110Hz (camera capture rate), placing limit on the projector (85Hz refresh rate) Projector latency accounts for 76% of total system latency! With occluder movement tolerance of 5cm, shadows are imperceptible up to 94 cm/sec (fast walking)
Matthew Flagg © Conclusion Presented new approach to VRP Occluder Detection in IR spectrum All processing moved to GPU 2 System Challenges Met Display Image Quality Shadow Perception Avoidance Shadows eliminated fast enough to accommodate walking
Matthew Flagg © Future Work Explore hardware solutions Recent results show an LCD projector having ½ the latency of a DLP and LCOS User Study VRP currently used in Collaborative Design Lab in School of Aerospace Engineering Replicate laboratory evaluation of passive VRP with new active VRP system Improve Image Quality Better seam blending