Research
Here are three recent projects:
ASAP -- Automated Shot Analysis Program
|
Shot
|
Start
Fr.
|
End
Fr.
|
In Time
|
Out
Time
|
Duration
|
Move
|
Clip
|
|
171
|
21873
|
21874
|
14:34:22
|
14:34:23
|
0:00:02
|
Hold
|
 |
|
|
|||||||
|
21875
|
21933
|
14:34:24
|
14:37:07
|
0:02:09
|
Tilt
up,
1.03819 Heights |
 |
|
|
|
|||||||
|
21934
|
21952
|
14:37:08
|
14:38:01
|
0:00:19
|
Hold
|
 |
We have devised a method for automated video indexing and scene characterization that meets the usage requirements of broadcast television and feature film editors. ASAP interprets source video material in a manner consistent with industry practice and provides output in a searchable form to aid in post production. It uses perspective estimation methods, originally developed for augmented reality applications, for cut detection and to parse camera movement, in conjunction with temporal filtering to resolve complex subject motion. We have conducted experiments applying ASAP to source footage from a professional music video and a complex section from a finished feature film, analysed with a rigorously-defined set of metrics. The results indicate that our techniques are robust and potentially effective for professional application.
Output from the
music video test can
be found here.
Output from the feature film test can
be found here.
Quincunx coding/BTPC
Binary Tree Predictive Coding (BTPC) is as an effective general-purpose image compression technique for both natural (photographic) and graphical pictures. It uses a non-causal, shape-adaptive predictor to decompose an image into a binary tree of prediction errors and zero blocks. Compression performance is comparable with JPEG for photographs, with GIF for graphics, and superior to the state of the art for composite images. An efficient (i.e. fast) implementation is available under the GNU public licence. BTPC has been integrated into commercial products, used as a front end for NN-based image classifiers, and favorably evaluated for medical image compression applications. We continue to make improvements to the basic algorithm, but most recent work has dealt with applying BTPC to video. Binary Tree Recursive Motion Estimation coding combines block-based and pel-recursive motion estimation in a binary-tree structure. It achieves a bit rate reduction of up to 20% for constant quality, over conventional block-based motion estimation and residue coding.
2D Object Based Coding
We have been investigating motion segmentation for MPEG-IV applications. The proposed new approach works top-down, first estimating a best fit motion prediction for the frame as a whole, then finding non-conforming regions, segmenting, and recalculating motion vectors for the independent segments. The procedure iterates to produce an accurate segmentation of the several moving objects in a video sequence.
As part of this work a new two-level picture compression scheme has been developed that performs better than all standard and known research methods.
Asymmetric binary tree partition image generated by Ahsan Shamim
