Video Retrieval by Learning Uncertainties in Concept Detection from Imbalanced Annotation Data
Video Retrieval by Learning Uncertainties in Concept Detection from Imbalanced Annotation Data
复制标题
DOI：
--
复制
发表时间：
2013
期刊：
影响因子：
--
通讯作者：
Kimiaki Shirahama;Kenji Kumabuchi;K. Uehara
中科院分区：
文献类型：
--
作者：
Kimiaki Shirahama;Kenji Kumabuchi;K. Uehara
- 基于概念的视频检索基于概念的检测结果来检索与查询相关的镜头，例如人、建筑物和汽车。然而，概念检测是“不确定的”，因为即使是最先进的方法也不能准确地检测各种概念。因此，我们介绍了一种视频检索方法，该方法使用“不确定性”对每个概念的检测中的不确定性进行建模。可解释性表示概念在镜头中存在（或不存在）的概率的上限。使用这种可验证性，可以有效地管理概念的假阳性和假阴性检测。我们通过估计注释了概念的存在和不存在的镜头之间的密度比来获得模糊度。然而，由于“不平衡问题”，注释随机采样的镜头不会导致适当的调整。这意味着，有概念的镜头数量通常比没有概念的镜头数量少得多。为了克服这一点，一个选择性的采样方法，优先采样未注释的镜头，这是类似的镜头已经注释的概念的存在。在TRECVID 2009视频数据上的实验结果验证了该方法的有效性。
—Concept-based video retrieval retrieves shots relevant to a query based on detection results of concepts, such as Person , Building and Car . However, concept detection is ‘uncertain’ because even state-of-the-art methods cannot accurately detect various concepts. Thus, we introduce a video retrieval method, which models the uncertainty in the detection of each concept using ‘plausibilities’. A plausibility represents an upper bound of probability that the concept is present (or absent) in a shot. Using such plausibilties, false positive and false negative detections of the concept can be effectively managed. We derive plausibilities by estimating the density ratio between shots annotated with the concept’s presence and absence. However, annotating randomly sampled shots does not lead appropriate plausibilities due to the ‘imbalanced problem’. This means that the number of shots where the concept is present is generally much smaller than the number of shots where it is absent. To overcome this, a selective sampling method is developed to preferentially sample unannotated shots, which are similar to shots already annotated with the concept’s presence. Experimental results on TRECVID 2009 video data validates the effectiveness of derived plausibilities.