Fast Object Segmentation in Unconstrained Video

Abstract

We present a technique for separating foreground objects from the background in a video. Our method is fast, fully automatic, and makes minimal assumptions about the video. This enables handling essentially unconstrained settings, including rapidly moving background, arbitrary object motion and appearance, and non-rigid deformations and articulations. In experiments on two datasets containing over 1400 video shots, our method outperforms a state-of-theart background subtraction technique [4] as well as methods based on clustering point tracks [6, 18, 19]. Moreover, it performs comparably to recent video object segmentation methods based on object proposals [14, 16, 27], while being orders of magnitude faster.

Efficient Image and Video Co-localization with Frank-Wolfe Algorithm

“…It is interesting to see that our performances are worse than [34]. In their method, they do not learn a model between the videos and use instead a novel unsupervised motion segmentation method.…”

“…Results are given in Table 4, where we compare to the co-localization method of [37], our image model with and without smoothing. Note that better results are obtained in [34] using unsupervised motion segmentation and appearance consistency within each video, which works particularly well for this dataset where objects of interest are moving. In contrast, our method focuses on trying to leverage appearance information across different videos in conjunction with temporal consistency.…”

Large Displacement Dynamic Scene Segmentation through Multiscale Saliency Flow

“…We associated each superpixel c i with a random variable l i ∈ L, each variable takes a value l i ∈ {0, 1} from the label set L ∈ {0, 1} N c . Inspired by similar works [12], [19], we define an objective function for the proposed MSF video segmentation as:…”

“…Our approach falls in this category as it also estimates motion cues across video sequences. Papazoglou et al [12] propose a fast object segmentation (FOS) algorithm which attempts to build dynamic appearance models of the object and background under the assumption that they change smoothly over time. An advantage of this approach is that it may be possible to handle spatio-temporal cues such as color and location in the labeling refinement stage.…”

“…On the other hand, the object-like regions lack motion interpretation and typically fail if the appearance or shape of foreground objects is changed. Trajectory cues in motion based approaches [6], [10], [12] can alleviate some of the problems inherent in spatial based approaches. However, the underlying assumption of motion based methods is that there exists relative motion between foreground and background, which cannot in general be assumed, as some parts or the whole body of objects may remain still for a number of frames.…”