Sampling Strategies for Real-Time Action Recognition

Feng, Shi; Petriu, Emil M.; Laganière, Robert

doi:10.1109/cvpr.2013.335

Cited by 92 publications

(72 citation statements)

References 23 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…They show that such a technique performs significantly better than the original STIP detectors. Shi et al [19] proposed that with proper sampling density, a state-of-the-art performance can be achieved by randomly discarding up to 92% of densely sampled interest points.…”

Section: Interest Point Selectionmentioning

confidence: 99%

“…The feature vectors in some of these channels can also be as high as 3-8 times than those used in original dense trajectories or our proposed ordered trajectories. Shi et al 2013 [19] suggested that random selection of 10,000 dense trajectories can yield 83.3% and 47.6%, respectively. By our proposed technique, we see that only less important information such as background is omitted, as shown visually in a few examples in Fig.…”

Section: Comparison With Other Techniquesmentioning

confidence: 99%

See 1 more Smart Citation

Ordered Trajectories for Large Scale Human Action Recognition

Murthy

Goecke

2013

2013 IEEE International Conference on Computer Vision Workshops

View full text Add to dashboard Cite

Recently, a video representation based on dense trajectories has been shown to outperform other human action recognition methods on several benchmark datasets. In dense trajectories, points are sampled at uniform intervals in space and time and then tracked using a dense optical flow field. The uniform sampling does not discriminate objects of interest from the background or other objects. Consequently, a lot of information is accumulated, which actually may not be useful. Sometimes, this unwanted information may bias the learning process if its content is much larger than the information of the principal object(s) of interest. This can especially escalate when more and more data is accumulated due to an increase in the number of action classes or the computation of dense trajectories at different scales in space and time, as in the Spatio-Temporal Pyramidal approach. In contrast, we propose a technique that selects only a few dense trajectories and then generates a new set of trajectories termed 'ordered trajectories'. We evaluate our technique on the complex benchmark HMDB51, UCF50 and UCF101 datasets containing 50 or more action classes and observe improved performance in terms of recognition rates and removal of background clutter at a lower computational cost.

show abstract

Section: Interest Point Selectionmentioning

confidence: 99%

Section: Comparison With Other Techniquesmentioning

confidence: 99%

Ordered Trajectories for Large Scale Human Action Recognition

Murthy

Goecke

2013

2013 IEEE International Conference on Computer Vision Workshops

View full text Add to dashboard Cite

show abstract

“…Their method however, uses low-level features for body-part estimation and hierarchical segmentation, and thus lacks robustness which limits their use for complex datasets. Most of these methods are predominantly global recognition methods and are not well-suited for use in the recognition of complex activities; however, methods like [12,30,36,44] that have performed relatively well on complex datasets have indirectly built coarse mid-level representations from low-level features.…”

Section: Low-level Descriptors For Activity Classificationmentioning

confidence: 99%

“…Introduced by [16], sparse space-time interest points and subsequent methods, such as local ternary patterns [41], joint sparse representations [12], dense interest points [37,30], better motion cues [14] and discriminative class-specific features [15], typically compute a bag-of-words representation out of local features and use them for classification. The work of [35] uses densely rather than sparsely sampled trajectories for better performance, and [36] builds upon this work to incorporate more types of low-level features while also accounting for camera motion.…”

Section: Low-level Descriptors For Activity Classificationmentioning

confidence: 99%

Part Bricolage: Flow-Assisted Part-Based Graphs for Detecting Activities in Videos

Shankar

Badrinarayanan

Cipolla

2014

Computer Vision – ECCV 2014

View full text Add to dashboard Cite

Abstract. Space-time detection of human activities in videos can significantly enhance visual search. To handle such tasks, while solely using low-level features has been found somewhat insufficient for complex datasets; mid-level features (like body parts) that are normally considered, are not robustly accounted for their inaccuracy. Moreover, the activity detection mechanisms do not constructively utilize the importance and trustworthiness of the features.This paper addresses these problems and introduces a unified formulation for robustly detecting activities in videos. Our first contribution is the formulation of the detection task as an undirected node-and edge-weighted graphical structure called Part Bricolage (PB), where the node weights represent the type of features along with their importance, and edge weights incorporate the probability of the features belonging to a known activity class, while also accounting for the trustworthiness of the features connecting the edge. Prize-Collecting-SteinerTree (PCST) problem [19] is solved for such a graph that gives the best connected subgraph comprising the activity of interest. Our second contribution is a novel technique for robust body part estimation, which uses two types of state-of-the-art pose detectors, and resolves the plausible detection ambiguities with pre-trained classifiers that predict the trustworthiness of the pose detectors. Our third contribution is the proposal of fusing the low-level descriptors with the mid-level ones, while maintaining the spatial structure between the features.For a quantitative evaluation of the detection power of PB, we run PB on Hollywood and MSR-Actions datasets and outperform the state-of-the-art by a significant margin for various detection paradigms.

show abstract

“…The comparative studies [19,21,23] are devoted to these different sampling methods, which have a great influence on the results. Hence, this step is also the first and key part of our work.…”

Section: Constructing Local Patch Vectorsmentioning

confidence: 99%

Local Patch Vectors Encoded by Fisher Vectors for Image Classification

et al. 2018

View full text Add to dashboard Cite

Abstract:The objective of this work is image classification, whose purpose is to group images into corresponding semantic categories. Four contributions are made as follows: (i) For computational simplicity and efficiency, we directly adopt raw image patch vectors as local descriptors encoded by Fisher vector (FV) subsequently; (ii) For obtaining representative local features within the FV encoding framework, we compare and analyze three typical sampling strategies: random sampling, saliency-based sampling and dense sampling; (iii) In order to embed both global and local spatial information into local features, we construct an improved spatial geometry structure which shows good performance; (iv) For reducing the storage and CPU costs of high dimensional vectors, we adopt a new feature selection method based on supervised mutual information (MI), which chooses features by an importance sorting algorithm. We report experimental results on dataset STL-10. It shows very promising performance with this simple and efficient framework compared to conventional methods.

show abstract

Sampling Strategies for Real-Time Action Recognition

Cited by 92 publications

References 23 publications

Ordered Trajectories for Large Scale Human Action Recognition

Ordered Trajectories for Large Scale Human Action Recognition

Part Bricolage: Flow-Assisted Part-Based Graphs for Detecting Activities in Videos

Local Patch Vectors Encoded by Fisher Vectors for Image Classification

Contact Info

Product

Resources

About