Better Exploiting Motion for Better Action Recognition

By Mihir Jain, Herv'e J'egou, Patrick Bouthemy. 2013

435 citations recorded by the source corpus.

Subfields: Computer Vision and Pattern Recognition

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

Citations in the corpus

Delving into Egocentric Actions (2015, positive): “The latter has been proved to be effective for action recognition [37, 12].”
Two-Stream Convolutional Networks for Action Recognition in Videos (2014, methodology): “The importance of camera motion compensation has been previously highlighted in [10, 26], where a global motion component was estimated and subtracted from the dense flow.”
Beyond short snippets: Deep networks for video classification (2015, methodology): “These approaches are able to aggregate local appearance and motion information using hand-crafted features such as Histogram of Oriented Gradients (HOG), Histogram of Optical Flow (HOF), Motion Boundary Histogram (MBH) around spatio-temporal interest points [17], in a dense grid [24] or around dense”
Beyond Gaussian Pyramid: Multi-skip Feature Stacking for action recognition (2014, neutral): “[7]’s approach incorporated a new motion descriptor.”
Actions ~ Transformations (2015, neutral): “STIP+BoW [23] 23.0% STIP+BoW [42] 43.9% DCS+VLAD [15] 52.1% CNN[21] 65.4% VLAD Encoding [56] 56.4% IDT+FV [51] 85.9% IDT+FV [51] 57.2% IDT+HSV [31] 87.9% Two Stream [39] 59.4% Two Stream [39] 88.0% IDT+HSV [31] 61.1% LSTM with 88.6% Two Stream [30] TDD+FV [52] 63.2% TDD+FV [52] 90.3% inputs…”
VideoLSTM convolves, attends and flows for action recognition (2016, background): “We reckon and value the impact of traditional video representations, e.g. , Wang and Schmid (2013), Jain et al. (2013), Sadanand and Corso (2012), Peng et al. (2016); 2014),Fernando et al. (2015) and Lan et al. (2015), and recent mixtures of shallow and deep encodings e.g. , Wang et al. (2015a) and…”
Human Action Recognition in Unconstrained Videos by Explicit Motion Modeling (2015, methodology): “Jain et al. [40] extended the work by considering the compensation of dominant motion in both tracking and encoding stages, which is different as Wang's work only used the MBH to consider the issue in the encoding stage.”
Learning Temporal Embeddings for Complex Video Analysis (2015, background): “Standard tasks in video such as classification and retrieval require a well engineered feature representation, with many proposed in the literature [1, 8, 13, 21, 25, 26, 27, 29, 32, 39, 40].”
Movement Pattern Histogram for Action Recognition and Retrieval (2014, methodology): “We estimate the background motion by computing homography between frames from the optical flow motion vectors (this is similar to [8] but instead of assuming affine motion between frames we use homography).”
Exploiting Image-trained CNN Architectures for Unconstrained Video Classification (2015, background): “This shortcoming has been addressed by extracting descriptors from spatiotemporal interest points ( e.g., [15, 4, 35, 16, 34]) or along estimated motion trajectories (e.g., [31, 11, 9, 32, 33]).”
A Review of Human Activity Recognition Methods (2015, methodology): “Jain et al. (2013) used motion compensation techniques to recognize atomic actions.”
Self-Paced Learning with Diversity (2014, neutral): “2013 [27] 62.”
Exploiting Feature and Class Relationships in Video Categorization with Regularized Deep Neural Networks (2015, result): “9 percent, outperforming many of the compared approaches [2], [11], [64], [65], [66], except a”
Sympathy for the Details: Dense Trajectories and Hybrid Classification Architectures for Action Recognition (2016, methodology): “There are several recent improvements to iDT, for instance, using motion compensation [5,38,39,6] and stacking of FVs to obtain a multi-layer encoding similar to mid-level representations [40].”
Hierarchical motion evolution for action recognition (2015, background): “The very popular features are dense trajectories [23, 25] and its variants [10, 8], which track densely sampled local interest points in the short term (e.”
Action Recognition by Hierarchical Mid-Level Action Elements (2015, neutral): “[14] 62.”
Optical flow modeling and computation: A survey (2015, background): “Action recognition is for instance an essential task for semantic interpretation of video content [1,130,256].”
Neural networks and neuroscience-inspired computer vision. (2014, background): “To date, many of the greatest successes in computer vision and object recognition have been with still images — this is perhaps not surprising given that we’ve only just now [73–75].”