Graphics Reference
In-Depth Information
Figure 6.12
High dynamic range video of a driving scene. Top row: input video with alternating short
and long exposures. Bottom row: high dynamic range video (tone-mapped). (From [Kang
et al. 03] c
2003 ACM, Inc. Included here by permission.) (See Color Plate IX.)
gorithm described in the “High Dynamic Range Video” paper are described as
follows.
1. Video capture.
The authors used a specialized digital camera to capture
the image sequence. The camera had a programmable control unit that allowed a
tethered computer to control the exposure in real time. The video is shot so that
dark and bright regions of the scene are exposed in rapid succession. The actual
range of exposures is adjusted dynamically to optimize effective capture. This
works much like the “auto-bracketing” feature available on many contemporary
digital still cameras: the camera determines the average scene brightness and then
captures a series of images with exposure times less than and greater than the
computed optimal exposure. In capturing video, these differently exposed images
are separate frames in the video sequence.
2. Frame registration.
Ultimately the captured images are collected into an
HDR video sequence, but this is complicated by the fact that the images are not
static. The camera used by the authors captured 15 frames per second, so there is
enough time between frames for the images to change appreciably. However, the
frame rate is fast enough that sequential frames are likely to be similar enough
to apply an automatic optical flow algorithm (see Chapter 5). However, the basic
search methods for pixel correspondence do not directly apply, because corre-
sponding pixel values vary widely between frames. At the same time, the differ-
ence in these values is needed to construct the HDR radiance map. The authors
describe an image correspondence method they call
HDR stitching
.
The HDR stitching algorithm works by constructing warped intermediate im-
ages between captured frames. This requires matching of corresponding pixels
