Webcam Synopsis: Peeking Around the World

Abstract

The

Summarizing visual data using bidirectional similarity

“…Note that most existing summarization algorithms (e.g., [1,14,16,3,15,10,12,8,17]) cannot exploit repetitiveness or redundancy of the visual data. This leads to significant scaling down or distortion of image patterns by these methods when the target size gets very small.…”

“…More sophisticated methods have been proposed for automatic retargeting by reorganizing the visual data (image or video) in a more compact way, while trying to preserve visual coherence of important (usually sparse) regions [3,8,10,11,12,14,15,16,17]. These methods can roughly be classified into three families: (i) Importancebased scaling methods [10,14,16] first identify important regions within the image (e.g., salient regions, faces, highmotion regions).…”

“…(ii) Importance-based cropping methods [3,11,15] provide nice results when the interesting information is concentrated in one region (spatial or temporal). (iii) Importance-based bin-packing meth-ods [8,12,14,17] further account for the main deficiency of cropping -the inability to capture spatially or temporally separated objects -by compact packing (spatial and/or temporal) of segmented important/salient regions/blobs.…”

Exploratory analysis of time-lapse imagery with fast subset PCA

“…In [14], Oh et al survey a series of video cataloging tools, collectively known as video abstraction, including various tools to select the few frames that summarize long portions of a video sequence. Summarizing long sections of video sequences has been done by detecting motion in spatio-temporal volumes and compositing different temporal segments of the video into a single image describing the overall motion in the image [15]. In each case, the authors focus on summarizing image sequences with high refresh rates.…”

VideoCut: Removing Irrelevant Frames by Discovering the Object of Interest

“…Nonetheless, our system exploits temporal information of videos in a novel way, which distinguishes itself from the image retrieval literature. In [16], activities in a video are condensed into a shorter period by simultaneously showing multiple activities. It does not intend to discover the frames that contain the user-desired OOI from limited user input.…”