[PDF] UCF101: A Dataset of 101 Human Actions Classes From Videos in The Wild | Semantic Scholar (2024)

A CNN–LSTM architecture where a pre-trained VGG16 convolutional neuronal networks extracts the features of the input video and a LSTM classifies the video in a particular class.

This work proposes an attention mechanism adapted to a CNN–LSTM base architecture that can be used for action recognition in videos and evaluates the performance of the system using accuracy as the evaluation metric.

I3D models considerably improve upon the state-of-the-art in action classification, reaching 80.2% on HMDB-51 and 97.9% on UCF-101 after pre-training on Kinetics, and a new Two-Stream Inflated 3D Conv net that is based on 2D ConvNet inflation is introduced.

A novel Cross-modality and Cross-action Modeling (CoCo) framework for ZSAR that significantly outperforms the state-of-the-art on three popular ZSAR benchmarks (i.e., Kinetics-ZSAR, UCF101 and HMDB51) under two different learning protocols in ZSAR.

A CNN–LSTM architecture is implemented that first, a pre-trained VGG16 convolutional neural network extracts the features of the input video, then an LSTM classifies the video in a particular class.

The dataset is described, the statistics are described, how it was collected, and some baseline performance figures for neural network architectures trained and tested for human action classification on this dataset are given.

The Action Transformer model for recognizing and localizing human actions in video clips is introduced and it is shown that by using high-resolution, person-specific, class-agnostic queries, the model spontaneously learns to track individual people and to pick up on semantic context from the actions of others.

TaiChi consists of unconstrained user-uploaded web videos containing camera motion and partial occlusions which pose new challenges to fine-grained action recognition compared to the existing datasets.

Experimental results on the UCF50, UCF101, and HMDB51 action datasets demonstrate that TS is comparable to state-of-the-arts, and outperforms many other methods; for HMDB the accuracy of 85.4%, compared to the best accuracy obtained by a deep method.

Different action recognition datasets are explored to highlight their ability to evaluate different models, and a usage is proposed for each dataset based on the content and format of data it includes, the number of classes and challenges it covers.

...

...

This paper uses the largest action video database to-date with 51 action categories, which in total contain around 7,000 manually annotated clips extracted from a variety of sources ranging from digitized movies to YouTube, to evaluate the performance of two representative computer vision systems for action recognition and explore the robustness of these methods under various conditions.

This paper presents a systematic framework for recognizing realistic actions from videos “in the wild”, and uses motion statistics to acquire stable motion features and clean static features, and PageRank is used to mine the most informative static features.

This paper construct video representations in terms of local space-time features and integrate such representations with SVM classification schemes for recognition and presents the presented results of action recognition.

This paper automatically discover relevant scene classes and their correlation with human actions, and shows how to learn selected scene classes from video without manual supervision and develops a joint framework for action and scene recognition and demonstrates improved recognition of both in natural video.

A new framework is proposed where actions are model actions using three dimensional occupancy grids, built from multiple viewpoints, in an exemplar-based HMM, where a 3D reconstruction is not required during the recognition phase, instead learned 3D exemplars are used to produce 2D image information that is compared to the observations.

A new framework is proposed where actions are model actions using three dimensional occupancy grids, built from multiple viewpoints, in an exemplar-based HMM, where a 3D reconstruction is not required during the recognition phase, instead learned 3D exemplars are used to produce 2D image information that is compared to the observations.

The method is fast, does not require video alignment and is applicable in many scenarios where the background is known, and the robustness of the method is demonstrated to partial occlusions, non-rigid deformations, significant changes in scale and viewpoint, high irregularities in the performance of an action and low quality video.

A framework for modeling motion by exploiting the temporal structure of the human activities, which represents activities as temporal compositions of motion segments, and shows that the algorithm performs better than other state of the art methods.

This paper generalizes the traditional MACH filter to video (3D spatiotemporal volume), and vector valued data, and analyzes the response of the filter in the frequency domain to avoid the high computational cost commonly incurred in template-based approaches.

A new method for detecting objects such as bags carried by pedestrians depicted in short video sequences by comparing the temporal templates against view-specific exemplars generated offline for unencumbered pedestrians, which yields a segmentation of carried objects using the MAP solution.

...

...