¡¡
The
ability to remove large objects in Images and videos is critical to many
applications, such as video editing and film post-production. Given an
input image or video, the goal is to remove the undesired objects and
reconstruct the corresponding unknown regions in the image or entire video
sequence based on the motion information. Based on our single image
completion method, we propose a novel approach to solve the video
completion problem, for the video sequences containing several planar
motion layers. Our method based on the assumptions that the overlapping
order of the motion layers in each frame is maintained the same throughout
the input videos, and there has no cross occlusion between the layers in
the video. For example, given that a video contains three layers 1, 2, and
3, if 1 occludes 2, and 2 occludes 3, 3 cannot occlude 1.
Based
on this assumption, we first apply a level set representation and graph cut
approach to achieve motion layer extraction. By exploiting the occlusion
order constraints on multiple consecutive frames, the occluded pixels and
the layer ordering are also explicitly determined. Then we remove the
undesired layer (the large object) and locate the corresponding unknown
areas in other layers for every frame. After selecting the reference frame,
we apply the motion compensation to partially or even fully fill the
unknown region in each layer. For the layers where some regions are still
missing, we develop a graph cut based region completion algorithm to
complete the missing data with the perceptually correct color-texture
information. Finally, based on the layer motion parameters, we project the
synthesized layers to render each new frame.
Related papers: [Click here]
Project
page: [Click
here]
In this
project, we present a real-time computer vision system for the monitoring
of the movement of dangerous actions at railroad intersections. The online
input video is processed in real time to detect the dangerous actions and a
set of alarms are triggered automatically.
We
designed our system based on a cost effective and portable principle. The
final deliverable system is powered by solar panels and is designed to be
water proof. It can be carried by standard sized pick up trucks and runs in
a rural area for long time.
Project page: [Click here]
Our
Computer Vision Group at University
of Central Florida
participated in two tasks in TRECVID 2004: High-Level Feature Extraction
and Story Segmentation. The part I involved in this project is "person
X" high level feature extraction.
Our
approach to find a specific person X combines text cues from the given
transcripts, a face detection method implemented based on a modified
version of the Haar-like feature face detector in OpenCV [1], and face
recognition based on a Support Vector Machine (SVM) classifier.
Related papers: [Click here]
