[Japanese | Thesis | Researches in Minoh Lab | Minoh Lab]
Technology of generating objects' image from an arbitrary viewpoint plays an important role in applications like displaying products in e-commerce. Such image can be generated when the object's 3D (three-dimensional) shape and its reflection property are acquired. As for 3D shape acquisition, laser range finders, the multi-baseline stereo method and the volume intersection technique are widely used.
As for reflection property acquisition, we can estimate the reflection property as lighting environment, the 3D shape of the object and the images from multiple viewpoints are given. Lighting environment consists of source of lights, their position and their power.
It is important to set up the lighting environment for the reflection property estimation. Camera images that observe reflected light are used for estimating the reflection property. If they have quantization error and don't exactly represent the reflected light, estimated reflection property contains error. Adequate lighting environment for relection property estimation reduces the error. Our goal is to search adequate lighting environment.
Variation in the power of incident light on the surface can result in variations in reflected light. The variation in reflected light increases quantization error of the images. Therefore, when we find the lighting environment which minimizes the variation in the power of incident light, our goal is achieved.
Power of incident light is calculated from the position and the power of the light and the position, normal vector of the surface. That means, the best lighting environment is decided according to the object's shape. In previous work, lights are arranged evenly on a sphere. On the other hand, we choose the most suitable lighting environment for the object shape.
Many point lights are arranged beforehand in order to achieve various appropriate lighting environments. We can change lighting environment by controlling the lights' ON/OFF switches.
Summary of our method is as follows. First, the visual hull is constructed by the volume intersection technique. It is expressed by sets of cubic lattices of a constant size (voxel). We consider the visual hull the shape of the object. Second, a voxel located on the surface of the visual hull is extracted as a surface voxel, and a normal vector of each surface voxel is calculated. Third, power of incident lights to each surface voxel is calculated from source of lights, their position and their power. Finally, we construct the lighting environment so that the variance of quantities of incident light to each surface voxel is minimized.
When the arranged number of the source lights increases, the number of combinations of ON/OFF becomes enormous. Therefore, calculating and comparing the decentralization by the combination of all ON/OFF is impossible in the range of practical calculation time. Then, to reduce the search space, we find the lights which ON/OFF setting can be automatically decided and exclude them from the search space.
We conducted simulation experiment and confirmed effectiveness of our approach. The shape of the object's body was acquired with cameras. Many point lights were evenly arranged. The variance of quantities of light environment decided by using our approach was smaller than the variance state to turn on all lights evenly arranged.