This project suggests that by using wavelet transform and vector quantization we can achieve pretty high compression ratio with minimun degradation in rendering quality.ġ Introduction: Light Fields, Wavelets and Vector Quantizationġ.1 Light Fields and Image Based Rendering In order to make light fields feasible even in low end computers we have to reduce the data size. However, the requirement for a huge amount of images causes problems both in mass storage and in memory management. Image based rendering based on light fields offers interactive frame rate that's unachievable by traditional geometry based rendering techniques.
You can download the full dataset here (24 GB).Light Field Compression using Wavelet Transform and Vector Quantization Chou, and Patrick Savill, “8i Voxelized Surface Light Field (8iVSLF) Dataset,” ISO/IEC JTC1/SC29 WG11 (MPEG) input document m42914, Ljubljana, July 2018. If you publish images of, or report performance results related to, these data, we request that you cite this document as: Maja Krivokuća, Philip A.
The terms of use of the dataset are governed by the License Agreement, which is an integral part of the dataset and which must accompany any copy of the dataset. For each point cloud in the 8iVSLF dataset, the attributes of an occupied voxel include: the red, green, and blue components of the surface colour as seen by each camera rig, and the x, y, z components of the voxel’s normal vector.Ĩi hereby makes available a new dataset of voxelized, high-resolution point clouds, as potential test material for MPEG standardization efforts, as well as for non-commercial use (subject to the accompanying license agreement) by the wider research community.
In each cube, only voxels that are near the surface of the subject are occupied. For the other point clouds in the 8iVSLF dataset, these occupy almost the entire set of 4096 voxels along their longest dimension (i.e., their height), so if we approximate their height as 1.8 m, we can say that a voxel in these datasets is approximately 0.44 mm on a side (i.e., 1.8 m / 4096 voxels ≈ 0.44 mm). Since the subject in this dataset takes up less than half the height of the 4096 x 4096 x 4096 cube of voxels, this makes her under 2 m tall, as expected. For this dataset, a voxel represents approximately 1 x 1 x 1 mm of the physical capture space. For the video sequence, the cube has been scaled so that it is the smallest bounding cube that contains the entire capture area. The inputs from all the clusters were then fused into a 3D surface.įor each of the contributed point clouds, a single spatial resolution is provided: a cube of 4096 x 4096 x 4096 voxels, known as depth 12 and denoted by vox12 in the name for each frame. Each cluster of cameras captured RGB and computed depth-from-stereo. The camera rigs were placed around the subject at approximately a couple of metres’ distance.
For the contributed video sequence, a 10-second period has been selected from the original captured sequence. For each point cloud in the 8iVSLF dataset, the full body of a human subject was captured by 39 synchronized RGB cameras configured in either 12 or 13 rigs, or clusters (each cluster acting as a logical RGBD camera), at 30 fps. The 8iVSLF dataset contains 1 high-resolution, 300-frame sequence, as well as 6 high-resolution single-frame point clouds. A dynamic voxelized point cloud is represented as a sequence of frames. A voxelized point cloud captured at one instant of time is a frame. Each occupied voxel may have attributes, such as colour, transparency, normals, curvature, and specularity. A voxel whose address is in the set is said to be occupied otherwise it is unoccupied. The coordinates may be interpreted as the address of a volumetric element, or voxel. Chou, and Patrick Savill).Ī voxelized point cloud is a set of points constrained to lie on a regular 3D grid, which, without loss of generality, may be assumed to be the integer lattice. Provided by 8i (Maja Krivokuća, Philip A.