Citation Context

...sentation with inadequate discriminative power for larger, more complex data sets. For example, on the 50-class UCF50 data set [1], the HOG/HOF method [18, 37] achieves 47.9% accuracy (as reported in =-=[17]-=-) whereas it achieves 85.6% on the smaller 9-class UCF Sports data set [30]. Other methods that seek a more semantically rich and discriminative Biking Javelin Jump Rope Fencing Bank of Action Detecto...

Citation Context

... videos) and a test set (884 videos). Training and test videos come from different movies. The performance is measured by mean average precision (mAP) over all classes, as in [22]. The HMDB51 dataset =-=[18]-=- is collected from a variety of sources ranging from digitized movies to YouTube videos. In total, there are 51 action categories and 6,766 video sequences. We follow the original protocol using three...

Citation Context

...orm poorly in improving recognition rates when applied to unconstrained videos; presumably due to misleading cues from multiple scene motions, low video quality etc. HMDB51. The recent HMDB51 dataset =-=[31]-=- contains 51 distinct action categories, each represented by at least 101 video clips. The total 6, 766 video clips were extracted from a wide range of varied sources. It is said to be one of the most...

Citation Context

... the art in action recognition is able to go beyond the toy problems of classifying simple human actions in controlled environment and considers the detection of actions in real movies or video clips =-=[11, 16]-=-. Despite these progresses, the existing descriptors suffer from an uncompleted handling of motion in the video sequence. Motion is arguably the most reliable source of information for action recognit...

Citation Context

...t the availability of large amounts of annotated image data by pretraining the spatial net on the ImageNet challenge dataset [1]. Our experiments on two challenging datasets (UCF-101 [20] and HMDB-51 =-=[15]-=-) show that the two video recognition streams are complementary, and our architecture significantly outperforms that of [13] and matches the state of the art [18, 22, 23], in spite of being trained on...

Citation Context

...asets KTH [31] cls 6 2,391 25 ≈200,000 160x120 USC gestures [21] cls 6 400 4 740x480 MSR action [43] cls,det 3 63 10 320x240 Movie datasets Hollywood2 [19] cls 12 1,707:69 UCF501 cls 50 >5,000 HMDB51 =-=[13]-=- cls 51 6,766 height:240 Coffee and Cigarettes [17] det 2 264:11 High Five [24] cls,det 4 300:23 Surveillance datasets PETS 2007 [11] det 3 10 32,107 768x576 UT interaction [28] cls,det 6 120 6 VIRAT ...

Citation Context

...t least part of the progress can be attributed to the development of more sophisticated low-level features. Currently, most successful methods are based on some form of local space-time features; see =-=[16, 43]-=- for recent evaluation studies. Most features are carefully engineered, while some recent work explores learning the low-level features from data [20, 45]. Once local features are extracted, often met...

Citation Context

...obal Reference Points Uncovering the global motion pattern in complex unconstrained videos is not an easy process. Typical solutions like foreground-background separation [15] and video stabilization =-=[28]-=- are very expensive to compute. We therefore pursue a simple solution by clustering the motion patterns of trajectories on the scene. The dominant pattern is treated as a reference point to characteri...

Citation Context

...listic event is a much more challenging task than just recognizing a set of motion discriminative actions (low level events) in standard datasets (such as KTH [1], UCF-Sports [2], UCF50 [3], and HMDB =-=[4]-=-). Some example video frames from complex event categories considered in this paper are shown in Fig. 1. Recently, the bag-of-words (BoW) approach has achieved impressive results in many recognition p...