Citation Context

...se according to their confidence scores. Only the selected regions are considered during optimization for efficiency. The objective of Eq.(1) is then efficiently optimized by dynamic programming (DP) =-=[6, 26]-=-. Note that using the p best tubes (p = 5 in all our experiments) for each video at each iteration (except the last one), instead of retaining only one candidate, increases the robustness of our appro...

Citation Context

...rance and geometric consistency. Assume two sets of region proposals have been extracted from vt and vu: Rt = R(vt) and Ru = R(vu). Let rt = (ft, lt) ∈ Rt be a region with its 8 × 8 HOG descriptor ft =-=[6, 13]-=- and its location lt, i.e., position and scale. The score for match m = (rt, ru) is decomposed into an appearance term ma = (ft, fu) and a geometry term mg = (lt, lu). Let x denote the location offset...

Citation Context

...ion. 1. Introduction Visual learning and interpretation is traditionally formulated as a supervised classification problem, with manually selected bounding boxes acting as (strong) supervisory signal =-=[7, 9]-=-. To reduce human effort and subjective biases in manual annotation, recent work has addressed the discovery and localization of objects from weakly-annotated or even unlabelled datasets [4, 5, 8, 26,...

Citation Context

...to detect object candidates. Similar approaches have been proposed for salient region detection [16], image cosegmentation [31, 32], and image colocalization [4]. Conventional object tracking methods =-=[35]-=- usually require annotations for at least one frame [12, 14, 34], or object detectors trained for target classes in a supervised manner [1, 2, 22]. Our method does not require such supervision and ins...

Citation Context

...ion. 1. Introduction Visual learning and interpretation is traditionally formulated as a supervised classification problem, with manually selected bounding boxes acting as (strong) supervisory signal =-=[7, 9]-=-. To reduce human effort and subjective biases in manual annotation, recent work has addressed the discovery and localization of objects from weakly-annotated or even unlabelled datasets [4, 5, 8, 26,...

Citation Context

...g tube for each video. We dub this iterative process a discovery and tracking procedure since finding the tubes maximizing foreground confidence across videos is akin to unsupervised object discovery =-=[4, 10, 11, 24, 27]-=-, whereas finding the tubes maximizing temporal consistency within a video is similar to object tracking [1, 2, 12, 22, 34, 35]. Interestingly, because we update the matching neighborhood structure at...

Citation Context

...pdate the neighborhood structure N by k nearest neighbor retrieval for each localized object region. At the first iteration, the nearest neighbor search is based on distances between GIST descriptors =-=[30]-=- of frames as the tube r is initialized as the entire video. From the second iteration, the metric is defined as the appearance similarity between potential object regions localized at the previous it...

Citation Context

.... Prest et al. [23] generate spatio-temporal tubes of object candidates, and select one of these per video through energy minimization. Since the candidate tubes rely only on clusters of point tracks =-=[3]-=-, this 1 Object Discovery across Videos Tracking Object within Video Figure 1. Given a noisy collection of videos, dominant objects are automatically localized as spatio-temporal tubes. The discovery ...

Citation Context

...n proposed for salient region detection [16], image cosegmentation [31, 32], and image colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame =-=[12, 14, 34]-=-, or object detectors trained for target classes in a supervised manner [1, 2, 22]. Our method does not require such supervision and instead alternates discovery and tracking of object candidates. The...

Citation Context

...ge colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame [12, 14, 34], or object detectors trained for target classes in a supervised manner =-=[1, 2, 22]-=-. Our method does not require such supervision and instead alternates discovery and tracking of object candidates. The problem we address is closely related to video object colocalization [15, 23], wh...

Citation Context

...ge colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame [12, 14, 34], or object detectors trained for target classes in a supervised manner =-=[1, 2, 22]-=-. Our method does not require such supervision and instead alternates discovery and tracking of object candidates. The problem we address is closely related to video object colocalization [15, 23], wh...

Citation Context

...n proposed for salient region detection [16], image cosegmentation [31, 32], and image colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame =-=[12, 14, 34]-=-, or object detectors trained for target classes in a supervised manner [1, 2, 22]. Our method does not require such supervision and instead alternates discovery and tracking of object candidates. The...

Citation Context

...g tube for each video. We dub this iterative process a discovery and tracking procedure since finding the tubes maximizing foreground confidence across videos is akin to unsupervised object discovery =-=[4, 10, 11, 24, 27]-=-, whereas finding the tubes maximizing temporal consistency within a video is similar to object tracking [1, 2, 12, 22, 34, 35]. Interestingly, because we update the matching neighborhood structure at...

Citation Context

...rance and geometric consistency. Assume two sets of region proposals have been extracted from vt and vu: Rt = R(vt) and Ru = R(vu). Let rt = (ft, lt) ∈ Rt be a region with its 8 × 8 HOG descriptor ft =-=[6, 13]-=- and its location lt, i.e., position and scale. The score for match m = (rt, ru) is decomposed into an appearance term ma = (ft, fu) and a geometry term mg = (lt, lu). Let x denote the location offset...

Citation Context

...tion or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used [3, 19, 20], while assuming that points from the same object have similar tracks. In =-=[17, 21]-=-, appearances of potential object and background are modeled and combined with motion information for the task. These methods produce results for individual videos and do not investigate relationships...

Citation Context

...y still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent research has started to handle the similar task in videos =-=[15, 23, 25, 33]-=-, and has shown that exploiting the space-time structure of the world, which is absent in static images, e.g., motion information, may be crucial for achieving object discovery or localization with le...

Citation Context

...signal [7, 9]. To reduce human effort and subjective biases in manual annotation, recent work has addressed the discovery and localization of objects from weakly-annotated or even unlabelled datasets =-=[4, 5, 8, 26, 28]-=-. However, this task is difficult and most approaches today still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent ...

Citation Context

...cts in our experiments of Section 5.3. Our setting is also related to object segmentation or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used =-=[3, 19, 20]-=-, while assuming that points from the same object have similar tracks. In [17, 21], appearances of potential object and background are modeled and combined with motion information for the task. These ...

Citation Context

...We consider a set of videos v, each consisting of T frames (images) vt (t = 1, . . . , T ), and denote by R(vt) a set of candidate regions identified in vt by some separate bottom-up proposal process =-=[18]-=-. Every region proposal is represented by a box in this paper. We also associate with vt a matching neighborhood N(vt) formed by the k closest frames wu among all videos w 6= v, according to a robust ...

Citation Context

...cts in our experiments of Section 5.3. Our setting is also related to object segmentation or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used =-=[3, 19, 20]-=-, while assuming that points from the same object have similar tracks. In [17, 21], appearances of potential object and background are modeled and combined with motion information for the task. These ...

Citation Context

...tion or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used [3, 19, 20], while assuming that points from the same object have similar tracks. In =-=[17, 21]-=-, appearances of potential object and background are modeled and combined with motion information for the task. These methods produce results for individual videos and do not investigate relationships...

Citation Context

... objects they contain. Video object cosegmentation aims to segment a detailed mask of common object out of videos. This problem has been addressed with weak supervision such as object class per video =-=[29]-=- and additional labels for a few frames that indicate whether the frames contain target object or not [33]. 1.2. Proposed approach We consider a set of videos v, each consisting of T frames (images) v...

Citation Context

...g tube for each video. We dub this iterative process a discovery and tracking procedure since finding the tubes maximizing foreground confidence across videos is akin to unsupervised object discovery =-=[4, 10, 11, 24, 27]-=-, whereas finding the tubes maximizing temporal consistency within a video is similar to object tracking [1, 2, 12, 22, 34, 35]. Interestingly, because we update the matching neighborhood structure at...

Citation Context

...signal [7, 9]. To reduce human effort and subjective biases in manual annotation, recent work has addressed the discovery and localization of objects from weakly-annotated or even unlabelled datasets =-=[4, 5, 8, 26, 28]-=-. However, this task is difficult and most approaches today still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent ...

Citation Context

...y still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent research has started to handle the similar task in videos =-=[15, 23, 25, 33]-=-, and has shown that exploiting the space-time structure of the world, which is absent in static images, e.g., motion information, may be crucial for achieving object discovery or localization with le...

Citation Context

... The discovery part establishes correspondences between frames across videos to detect object candidates. Similar approaches have been proposed for salient region detection [16], image cosegmentation =-=[31, 32]-=-, and image colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame [12, 14, 34], or object detectors trained for target classes in a supervise...

Citation Context

...k supervision such as an object label per video [33] and additional labels for a few frames that indicate whether the frames contain a target object or not [38]. Finally, spatio-temporal proposals of =-=[16, 24]-=- and action localization [39, 42] are relevant to our work as they also return spatio-temporal tubes as output. However, our method localizes an object through a single volume, whereas the proposals [...

Citation Context

...abel per video [33] and additional labels for a few frames that indicate whether the frames contain a target object or not [38]. Finally, spatio-temporal proposals of [16, 24] and action localization =-=[39, 42]-=- are relevant to our work as they also return spatio-temporal tubes as output. However, our method localizes an object through a single volume, whereas the proposals [16, 24] form a large number of hy...

Citation Context

...signal [7, 9]. To reduce human effort and subjective biases in manual annotation, recent work has addressed the discovery and localization of objects from weakly-annotated or even unlabelled datasets =-=[4, 5, 8, 26, 28]-=-. However, this task is difficult and most approaches today still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent ...

Citation Context

...gmentation or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used [3, 22, 23], assuming that points from the same object have similar tracks. In =-=[19, 20, 25, 35]-=-, the appearance of potential object and background regions is modeled and combined with motion information. These methods produce results for individual videos and do not investigate relationships be...

Citation Context

...k supervision such as an object label per video [33] and additional labels for a few frames that indicate whether the frames contain a target object or not [38]. Finally, spatio-temporal proposals of =-=[16, 24]-=- and action localization [39, 42] are relevant to our work as they also return spatio-temporal tubes as output. However, our method localizes an object through a single volume, whereas the proposals [...

Citation Context

...n proposed for salient region detection [16], image cosegmentation [31, 32], and image colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame =-=[12, 14, 34]-=-, or object detectors trained for target classes in a supervised manner [1, 2, 22]. Our method does not require such supervision and instead alternates discovery and tracking of object candidates. The...

Citation Context

...y still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent research has started to handle the similar task in videos =-=[15, 23, 25, 33]-=-, and has shown that exploiting the space-time structure of the world, which is absent in static images, e.g., motion information, may be crucial for achieving object discovery or localization with le...

Citation Context

...abel per video [33] and additional labels for a few frames that indicate whether the frames contain a target object or not [38]. Finally, spatio-temporal proposals of [16, 24] and action localization =-=[39, 42]-=- are relevant to our work as they also return spatio-temporal tubes as output. However, our method localizes an object through a single volume, whereas the proposals [16, 24] form a large number of hy...

Citation Context

...y still lay significantly behind strongly-supervised methods. With the ever growing popularity of video sharing sites such as YouTube, recent research has started to handle the similar task in videos =-=[15, 23, 25, 33]-=-, and has shown that exploiting the space-time structure of the world, which is absent in static images, e.g., motion information, may be crucial for achieving object discovery or localization with le...

Citation Context

... The discovery part establishes correspondences between frames across videos to detect object candidates. Similar approaches have been proposed for salient region detection [16], image cosegmentation =-=[31, 32]-=-, and image colocalization [4]. Conventional object tracking methods [35] usually require annotations for at least one frame [12, 14, 34], or object detectors trained for target classes in a supervise...

Citation Context

...gmentation or cosegmentation in videos. For video object segmentation, clusters of long-term point tracks have been used [3, 22, 23], assuming that points from the same object have similar tracks. In =-=[19, 20, 25, 35]-=-, the appearance of potential object and background regions is modeled and combined with motion information. These methods produce results for individual videos and do not investigate relationships be...

Citation Context

...riments) for each video at each iteration except the last one, instead of retaining only one candidate at each iteration, increases the robustness of our approach. This agrees with the conclusions of =-=[4]-=- in the still image domain, and has also been confirmed empirically by our experiments. We obtain p best tubes by sequential DPs, which iteratively remove the best tube and re-run DP again.1 5. Implem...