Research Projects | NTT-MIT Research Collaboration |
Image Based Synthetic Aperture RenderingMIT9904-14 Start date: 07/99 |
Leonard McMillan and Julie
Dorsey MIT LCS Hiroshi Murase NTT |
Project summary |
Project description |
Our representation uses a two-dimensional array of images to form a four-dimensional light field of any given scene. From this light field, we can compute novel new views of the captured scene with greater flexibility than a real camera and higher quality than a three-dimensional computer graphics model. Our work addresses the following problems:
The ultimate goal of this project is to construct a randomly accessible two-dimensional camera array capable of synthesizing novel views of dynamic scenes. |
Research Areas |
We are designing a two-dimensional camera array for acquiring and processing light fields in real time. In fact, the entire camera array can be regarded as a dynamic light field data structure. The pixels of each camera in the array are randomly accessible by the host computer system, where the array is mapped as a block of memory. These dynamic light fields are accessed "on-the-fly" to generate the desired images. The memory-like interface used by our camera array significantly reduces latency in the image generation process and makes the most effective use of available bandwidth. The camera array will employ a modular design consisting of a motherboard, sensor pods that support a range of imagers, and a PCI interface to the host, optimized for high throughput. |
We are also developing new processing techniques for light fields. First, we have added the ability to vary the apparent focus and depth-of-field within a light field using intuitive camera-like controls such as a variable aperture and focus ring. However, unlike a tradition camera, we allow for more general and flexible focal surfaces than the typical focal plane. Our techniques are based on the dynamic reparameterization of the light-field data structure in order to optimize and exhibit greater control over the image reconstruction process. Our reparameterization methods operate independent of scene geometry; we do not need to recover actual or approximate geometry of the scene for focusing. Our techniques also allow for the use of multiple focal surfaces within a single image rendering. |
|
We are also developing low-cost devices for acquiring light fields. Not only does this effort provide direct benefits for our research, but it also provides many new opportunities. The small size and portability of this system allows us to easily acquire outdoor light fields of natural scenes, and it is also currently our fastest acquisition device. In order to minimize cost we chose to use an off-the-shelf flatbed scanner as the digitizing element, in combination with a plastic lens array. However, intrinsic properties of scanner limit their use for acquiring images. We have developed a combination of low-cost optics and correction software to overcome these limitations. Currently we compensate for the following: color balance, radial distortion, camera calibration, noise removal, and sub-image segmentation.
|
We are also working on passive autostereoscopic devices for the direct viewing of light fields. Our display allows multiple viewers to see a light field stereoscopically from a wide viewing range and it requires no tracking. The key component of our viewing system is a prefabricated lens-array sheet with a thickness equal to the focal length of each lenslet. This allows us to make a viewer-dependent image with resolution equal to the size of each lenslet in the array. Currently, we are using a lens array in a hex pattern, where each lens has a diameter of approx. 0.1" and a focal length of 0.12". Under the entire sheet, we place a large composite image of smaller hexagonal image. Each small image is a picture of a light field taken from some point in space. When viewed under the lenses, a different pixel in the small image is seen for each view direction. In addition, each lens performs interpolation, due to the lens optics. When the entire array is viewed, each lens takes on the role of a single pixel in the final autostereoscopic image. |
Demos, movies and other examples |
Dynamically Reparameterized Light Fields Animations
|
These animation show how we can freely move around our dynamically reparameterized light fields. Throughout the MovingToys animation, we change the focal plane, camera orientation, and camera position. In the MovingTree animation, we vary the position and orientation of the camera. Because there is only one focal plane, a considerable amount of ghosting appears in these animations. This ghosting can be eliminated by 1) using a focal surface that conforms to the geometry of the scene, 2) using multiple focal planes that pass through the key features of the scene, 3) using a larger aperture such that ghosting turns into blur, or 4) sampling the camera surface at a higher resolution (using more cameras). The black regions occur when there is no ray in available in the ray database. |
|
The principal investigators |
Presentations and posters |
Publications |
Proposals and progress reports |
NTT Bi-Annual Progress Report, July to December 1999:
For more information |
To go here: links to the AI Lab, LCS, the main NTT-MIT page, and some buttons for: