没有合适的资源?快使用搜索试试~ 我知道了~
从密集的颜色和稀疏的交互式基于点的建模...doc
1.该资源内容由用户上传,如若侵权请联系客服进行举报
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
2.虚拟产品一经售出概不退款(资源遇到问题,请及时私信上传者)
版权申诉
0 下载量 44 浏览量
2024-04-02
14:31:40
上传
评论
收藏 11.4MB DOC 举报
温馨提示
试读
11页
从密集的颜色和稀疏的交互式基于点的建模...doc
资源推荐
资源详情
资源评论
1001 Acquisition Viewpoints
Efficient and Versatile View-Dependent Modeling of Real-World Scenes
Figure 1: Acquisition system
Figure 2: Typical acquisition path with 3,684 viewpoints
Figure 3: Images rendered from plant model acquired in 10 minutes.
Abstract
Three dimensional modeling is a severe bottleneck for computer graphics applications. Manual modeling is time
consuming and, even so, the resulting models fail to capture the true complexity of real world scenes. Automated modeling
based on acquiring color and depth data is a promising alternative. The conventional approach is to sample the scene
densely from a sparse set of acquisition viewpoints. Unfortunately, even with careful view planning, a sparse set of
acquisition viewpoints does not and cannot ensure adequate coverage for complex scenes. Moreover, the approach is
inefficient due to the considerable data redundancy between neighboring acquisition viewpoints: acquisition from an
additional viewpoint has the same cost while it contributes fewer and fewer new samples.
We propose an automated modeling approach based on sampling the scene sparsely from a dense set of acquisition
viewpoints. We show that the sparse data quickly accumulates to generate models with good scene coverage. The sparse
depth is acquired efficiently and robustly, which enables an interactive, operator-in-the-loop acquisition pipeline. We
describe a modeling system that implements this approach. The system acquires scenes with complex geometry and complex
reflective properties from thousands of viewpoints in minutes. The resulting model has a compact memory footprint and it
supports photorealistic rendering at interactive rates. The system is robust, yet it does not require displacing scene objects or
altering scene lighting conditions.
Categories and Subject Descriptors (ACM CCS): I.3.3. [Computer Graphics]—Three-Dimensional Graphics and Realism.
submitted to EUROGRAPHICS 2008.
2
1. Introduction
Constructing high-fidelity 3D models of real-world
scenes is an important bottleneck for many computer
graphics applications. The conventional approach of
manual modeling using CAD or animation software
requires artistic talent and a huge time investment.
Automated modeling based on acquiring color and depth
data is a promising alternative.
The typical automated modeling approach is to densely
sample the color and geometry of the scene from several
acquisition viewpoints, and then to merge the datasets to
obtain the scene model. However, dense depth sampling of
complex scenes remains a challenging process that can take
tens of minutes per acquisition viewpoint (e.g. sequential
scanning in laser rangefinding, robust off-line
correspondence computation in depth from stereo, or
sequential light pattern projection in depth from structured
light). This limits acquisition to a few viewpoints.
Unfortunately, even with optimal acquisition viewpoint
planning, complex scenes cannot be adequately sampled
from only a sparse set of viewpoints.
Instead of dense depth sampling from a sparse set of
acquisition viewpoints we propose an automated modeling
approach based on sparse depth sampling from a dense set
of acquisition viewpoints. We show that the sparse per-
viewpoint depth data quickly accumulates to generate
models with superior scene coverage (see Section 4.3).
Moreover, sparse depth can be acquired robustly and
efficiently, which enables an interactive, operator-in-the-
loop, acquisition pipeline. The operator detects and
addresses acquisition problems right away, and aims the
acquisition device such as to maximize the impact of the
sparse data, which ensures that a quality model is obtained
in a single scanning session.
The sparse-depth/dense-viewpoint (SDDV) approach
enables adding detail at linear cost: a quick scan can be
refined at will by revisiting scanned regions in order to
increase model fidelity. This is in sharp contrast with the
dense-depth/sparse-viewpoint (DDSV) approach where,
due to data redundancy, acquisition from each additional
viewpoint comes at the same cost yet contributes fewer and
fewer new samples.
Another important advantage of the SDDV modeling
approach is an increased robustness of the acquisition of
reflective surfaces. Such surfaces are challenging for
passive methods (e.g. stereo) because they complicate
correspondence searching, and also for active methods (e.g.
laser rangefinding, structured light) because they limit the
amount of emitted energy reflected back to the sensor. Our
approach of interactive acquisition from a dense set of
viewpoints allows the operator to avoid grazing scanning
angles.
We describe a modeling system that implements the
SDDV modeling approach (Figure 1). The system acquires
the scene from thousands of viewpoints in minutes (Figure
2). The color and depth data is combined into a view
dependent model, which produces quality novel views of
the scene at interactive rates (Figure 3, Figure 10, and
accompanying video). The system is versatile—it supports
complex geometry and surfaces with complex reflective
properties—and robust—all the models shown here were
acquired in a single take in one afternoon. We currently
target scenes that fit in a 0.5m cube. However, the system
has ample depth acquisition range, the resulting models are
compact, and the system does not require altering scene
lighting conditions or displacing scene objects, which are
important pre-conditions for a future extension to SDDV
modeling of large environments.
2. Prior work
We give a brief discussion of prior work organized
according to the method used for depth acquisition.
No-depth methods
One automated modeling approach is to bypass depth
acquisition altogether and to sample the scene color densely
from a dense set of acquisition viewpoints. The resulting
4D ray database (light field [LH96], and lumigraph
[GGS*96]) supports photorealistic rendering at interactive
rates. Light fields and their extensions remain the only
approach for acquiring extremely complex scenes (e.g.
feathers, fur, translucent gaze [MPN*02]). However, the
approach suffers from the disadvantages of large model
sizes. Panoramas [Che95] and their extensions are compact
2D ray databases, but the reduction in size comes at the
price of restricting the desired viewpoint to the center of the
panorama. All ray database methods preclude quantitative
applications that require explicit geometry.
User-specified-depth methods
Another approach is to rely on the user to specify a
coarse geometric model of the scene using an interactive
software tool that leverages model and image space
geometric constraints [DTM96, HH02, QTZ*06]. The
approach takes advantage of the user’s understanding of the
scene, who specifies the most relevant geometric entities.
However, the method scales poorly with the geometric
complexity of the scene, when manual geometry
specification becomes prohibitively expensive.
Dense-depth methods
One of the oldest and most appealing ways of acquiring
scene geometry is depth from stereo, where the disparity of
corresponding points in overlapping photographs is
translated into depth [PG02, DRR03, LHY*05]. A simple
camera suffices to capture scenes on small or large scale,
indoors or outdoors. However, finding correspondences is a
challenging problem, relying on the presence of color
texture.
.
submitted to EUROGRAPHICS 2008
3
In depth from structured light one camera of the stereo
pair is replaced with a light source that casts a special light
pattern in the field of view of the remaining camera. The
approach simplifies the search for correspondences and is
robust to surfaces that lack color texture. Using complex
patterns of light enables recovering dense depth maps from
a few frames [RHL02, KGG03], but the approach relies on
video projectors which are bulky and require strict control
of scene lighting.
The alternative is to use a simpler and brighter light
pattern such as a plane of laser light. Such triangulation
laser rangefinding produces high resolution depth maps
[Lev00, GFS05] but at a large time cost due to sequential
scanning. Moreover, the laser stripe degrades quickly with
the distance to the acquired surface, which requires that the
acquisition device be kept close to the scanned surface or to
use laser source power levels that are not eye safe.
Complex geometry is also a challenge for the laser stripe
approach due to laser scattering and secondary reflections.
A third approach for dense depth acquisition is time-of-
flight laser rangefinding [MNP99, STY03, GFS05]. Like in
the case of triangulation laser rangefinding, quality depth
maps are acquired but only at a substantial time cost.
However, the approach allows scanning from afar. Both
laser rangefinding approaches do not acquire color directly.
One disadvantage common to all dense depth methods is
the high redundancy between depth maps. Another
disadvantage is slow acquisition which limits the number of
acquisition locations, which in turn leads to poor scene
coverage.
Interactive methods
We are not the first to notice the advantages of an
interactive automated modeling pipeline. An early type of
interactive modeling system allows the operator to move a
light pattern inside the field of view of the fixed camera
[Bor*98, BP99]. This enables the operator to avoid grazing
angles between the scene surfaces and the pattern of light,
but is fundamentally limited to a single viewpoint by the
fixed camera. The most the operator can do is to assign
depth to all the pixels of the camera image, which is
equivalent to one dense depth map.
The single-viewpoint limitation is overcome by systems
that employ a laser stripe triangulation rangefinder whose
position and orientation is provided by an electromagnetic
[FFG*96] or mechanical [Hi00] tracker. However, the laser
stripe limitations discussed above remain. A different
approach for achieving interactive acquisition from many
viewpoints is to move the targeted scene in the field of
view of a fixed scanner using a rotating platter [TOF05], or
by relying on the operator to rotate the object [RHL02].
Such systems are essentially object scanners and scale
poorly with the size of the scene.
The ModelCamera system [PSB03, BPM06] employs a
structured light pattern that consists of a matrix of laser
beams, which ensures robust acquisition of sparse depth
data at interactive rates. Our acquisition device is
similar to the ModelCamera and we describe the
similarities and differences in Section 4.1. The
ModelCamera scans simple scenes (e.g. large pieces of
furniture) in a hand-held mode, without reliance on an
external tracker [PSB03]. Complex scenes (e.g. plants,
rooms, buildings) are acquired by panning and tilting the
ModelCamera about a fixed acquisition viewpoint
[BPM06]. The single viewpoint constraint limits the
fidelity of the acquired models.
3. System overview
We developed a system that executes the sparse-
depth/dense-viewpoint automated modeling approach. We
employ a hand-held acquisition device that acquires dense
color and sparse depth at interactive rates (Section 4.1).
The operator sweeps the scene with the device and
monitors the acquisition process through immediate visual
feedback provided on a nearby monitor (Figure 1). Data
acquisition proceeds in two steps: depth acquisition
(Section 4.2), which achieves adequate scene coverage
(Section 6.1), followed by color acquisition (Section 4.3).
The depth and color data is combined into a view-
dependent model for rendering (Section 0).
4. Acquisition
To support the SDDV approach the system should:
- acquire sparse depth efficiently and robustly
- acquire high-quality color
- allow the operator to freely position the acquisition device
- provide real-time feedback during acquisition.
4.1. Acquisition device
We employ an acquisition device similar to the
ModelCamera [BPM06]. Like the ModelCamera, our
acquisition device consists of a video camera with an
attached laser system that casts a matrix of laser beams into
the field of view of the camera (Figure 4). Each beam
creates a bright dot in the frame which is detected and
triangulated into a depth sample. By construction, the laser
beams project to disjoint image plane epipolar segments
which makes dot detection efficient and robust. Compared
to a laser stripe, the dot pattern has the advantage of higher
brightness and of fewer scattering and secondary reflection
artifacts. The baseline of 20 cm provides a depth range of
50 to 400 cm.
Whereas the ModelCamera was restricted to a single
acquisition viewpoint using a parallax-free pan-tilt bracket,
our acquisition device is attached to a mechanical tracking
arm which allows 6-degree-of-freedom motion. The
operator can position and orient the acquisition device
freely within a sphere with a radius of 60 cm to acquire
剩余10页未读,继续阅读
资源评论
百态老人
- 粉丝: 2129
- 资源: 2万+
下载权益
C知道特权
VIP文章
课程特权
开通VIP
上传资源 快速赚钱
- 我的内容管理 展开
- 我的资源 快来上传第一个资源
- 我的收益 登录查看自己的收益
- 我的积分 登录查看自己的积分
- 我的C币 登录后查看C币余额
- 我的收藏
- 我的下载
- 下载帮助
安全验证
文档复制为VIP权益,开通VIP直接复制
信息提交成功