Efficient manual annotation of videos by in-painting of previously detected objects
Publication Date: 2015-Jun-23
The IP.com Prior Art Database
We propose a method for reducing clutter in video for the purpose of efficient manual annotation (labeling) of a video sequence. The method consists of 3 steps: Object detection and tracking, image in-painting of the areas corresponding to detected video tubes, and highlighting of objects detected with low confidence. The adjsuted video is then presented to the human operator who is then able to efficiently annotate the remaning objects, as the video is much less cluttered.
Page 01 of 2
Efficient manual annotation of videos by in - -painting of previously detected objects
painting of previously detected objects
This paper presents a system for reducing clutter in video for the purpose of efficient manual annotation (labeling) of a video sequence.
Step 1: Automatically or manually annotate video sequence.
E.g. detect existence and position of cars and people and other objects of relevance using any method known in literature. Optionally track these objects along the video sequence.
Step 2: The objects detected with high confidence are removed from the video using in-painting methods. In the resulting video any remaining undetected objects now appear more salient to the human viewer and thus can be more easily detected.
Step 3 (Optional): Areas that are detected with low confidence are not removed but may be highlighted such that the attention of the human reviewers (of the video sequence) can be brought to these areas.
In addition: If one has access to many recordings of the same scene (either from the same video or from another recording) that information can be used to model the background and improve the results of in-painting process.
Page 02 of 2