Action Classification In Videos Research Articles

XKD: Cross-Modal Knowledge Distillation with Domain Alignment for Video Representation Learning

We present XKD, a novel self-supervised framework to learn meaningful representations from unlabelled videos. XKD is trained with two pseudo objectives. First, masked data reconstruction is performed to learn modality-specific representations from audio and visual streams. Next, self-supervised cross-modal knowledge distillation is performed between the two modalities through a teacher-student setup to learn complementary information. We introduce a novel domain alignment strategy to tackle domain discrepancy between audio and visual modalities enabling effective cross-modal knowledge distillation. Additionally, to develop a general-purpose network capable of handling both audio and visual streams, modality-agnostic variants of XKD are introduced, which use the same pretrained backbone for different audio and visual tasks. Our proposed cross-modal knowledge distillation improves video action classification by 8% to 14% on UCF101, HMDB51, and Kinetics400. Additionally, XKD improves multimodal action classification by 5.5% on Kinetics-Sound. XKD shows state-of-the-art performance in sound classification on ESC50, achieving top-1 accuracy of 96.5%.

Mutual modality learning for video action classification

The construction of models for video action classification progresses rapidly. However, the performance of those models can still be easily improved by ensembling with the same models trained on different modalities (e.g. Optical flow). Unfortunately, it is computationally expensive to use several modalities during inference. Recent works examine the ways to integrate advantages of multi-modality into a single RGB-model. Yet, there is still room for improvement. In this paper, we explore various methods to embed the ensemble power into a single model. We show that proper initialization, as well as mutual modality learning, enhances single-modality models. As a result, we achieve state-of-the-art results in the Something-Something-v2 benchmark.

Video Action Recognition by Combining Spatial-Temporal Cues with Graph Convolutional Networks

Video action recognition relies heavily on the way spatio-temporal cues are combined in order to enhance recognition accuracy. This issue can be addressed with explicit modeling of interactions among objects within or between videos, such as the graph neural network, which has been shown to accurately model and represent complicated spatial- temporal object relations for video action classification. However, the visual objects in the video are diversified, whereas the nodes in the graphs are fixed. This may result in information overload or loss if the visual objects are too redundant or insufficient for graph construction. Segment level graph convolutional networks (SLGCNs) are proposed as a method for recognizing actions in videos. The SLGCN consists of a segment-level spatial graph and a segment-level temporal graph, both of which are capable of simultaneously processing spatial and temporal information. Specifically, the segment-level spatial graph and the segment-level temporal graph are constructed using 2D and 3D CNNs to extract appearance and motion features from video segments. Graph convolutions are applied in order to obtain informative segment-level spatial-temporal features. A variety of challenging video datasets, such as EPIC-Kitchens, FCVID, HMDB51 and UCF101, are used to evaluate our method. In experiments, it is demonstrated that the SLGCN can achieve performance comparable to the state-of-the-art models in terms of obtaining spatial-temporal features.

Class-Difficulty Based Methods for Long-Tailed Visual Recognition

Long-tailed datasets are very frequently encountered in real-world use cases where few classes or categories (known as majority or head classes) have higher number of data samples compared to the other classes (known as minority or tail classes). Training deep neural networks on such datasets gives results biased towards the head classes. So far, researchers have come up with multiple weighted loss and data re-sampling techniques in efforts to reduce the bias. However, most of such techniques assume that the tail classes are always the most difficult classes to learn and therefore need more weightage or attention. Here, we argue that the assumption might not always hold true. Therefore, we propose a novel approach to dynamically measure the instantaneous difficulty of each class during the training phase of the model. Further, we use the difficulty measures of each class to design a novel weighted loss technique called ‘class-wise difficulty based weighted (CDB-W) loss’ and a novel data sampling technique called ‘class-wise difficulty based sampling (CDB-S)’. To verify the wide-scale usability of our CDB methods, we conducted extensive experiments on multiple tasks such as image classification, object detection, instance segmentation and video-action classification. Results verified that CDB-W loss and CDB-S could achieve state-of-the-art results on many class-imbalanced datasets such as ImageNet-LT, LVIS and EGTEA, that resemble real-world use cases.

Co-segmentation inspired attention module for video-based computer vision tasks

Video-based computer vision tasks can benefit from estimation of the salient regions and interactions between those regions. Traditionally, this has been done by identifying the object regions in the images by utilizing pre-trained models to perform object detection, object segmentation and/or object pose estimation. Although using pre-trained models is a viable approach, it has several limitations in the need for an exhaustive annotation of object categories, a possible domain gap between datasets and a bias that is typically present in pre-trained models. In this work, we propose to utilize the common rationale that a sequence of video frames capture a set of common objects and interactions between them, thus a notion of co-segmentation between the video frame features may equip the model with the ability to automatically focus on task-specific salient regions and improve the underlying task’s performance in an end-to-end manner. In this regard, we propose a generic module called “Co-Segmentation inspired Attention Module” (COSAM) that can be plugged in to any CNN model to promote the notion of co-segmentation based attention among a sequence of video frame features. We show the application of COSAM in three video-based tasks namely: (1) Video-based person re-ID, (2) Video captioning, & (3) Video action classification and demonstrate that COSAM is able to capture the task-specific salient regions in video frames, thus leading to notable performance improvements along with interpretable attention maps for a variety of video-based vision tasks, with possible application to other video-based vision tasks as well.

Light-Weight Classification of Human Actions in Video with Skeleton-Based Features

An approach to human action classification in videos is presented, based on knowledge-aware initial features extracted from human skeleton data and on further processing by convolutional networks. The proposed smart tracking of skeleton joints, approximation of missing joints and normalization of skeleton data are important steps of feature extraction. Three neural network models—based on LSTM, Transformer and CNN—are developed and experimentally verified. The models are trained and tested on the well-known NTU-RGB+D (Shahroudy et al., 2016) dataset in the cross-view mode. The obtained results show a competitive performance with other SOTA methods and verify the efficiency of proposed feature engineering. The network has a five times lower number of trainable parameters than other proposed methods to reach nearly similar performance and twenty times lower number than the currently best performing solutions. Thanks to the lightness of the classifier, the solution only requires relatively small computational resources.

Sound Can Help Us See More Clearly.

In the field of video action classification, existing network frameworks often only use video frames as input. When the object involved in the action does not appear in a prominent position in the video frame, the network cannot accurately classify it. We introduce a new neural network structure that uses sound to assist in processing such tasks. The original sound wave is converted into sound texture as the input of the network. Furthermore, in order to use the rich modal information (images and sound) in the video, we designed and used a two-stream frame. In this work, we assume that sound data can be used to solve motion recognition tasks. To demonstrate this, we designed a neural network based on sound texture to perform video action classification tasks. Then, we fuse this network with a deep neural network that uses continuous video frames to construct a two-stream network, which is called A-IN. Finally, in the kinetics dataset, we use our proposed A-IN to compare with the image-only network. The experimental results show that the recognition accuracy of the two-stream neural network model with uesed sound data features is increased by 7.6% compared with the network using video frames. This proves that the rational use of the rich information in the video can improve the classification effect.

Multiperson interaction recognition in images: A body keypoint based feature image analysis

AbstractMost interaction recognition approaches have been limited to single‐person action classification in videos. However, for still images where motion information is not available, the task becomes more complex. Aiming to this point, we propose an approach for multiperson human interaction recognition in images with keypoint‐based feature image analysis. Proposed method is a three‐stage framework. In the first stage, we propose feature‐based neural network (FCNN) for action recognition trained with feature images. Feature images are body features, that is, effective distances between a set of body part pairs and angular relation between body part triplets, rearranged in 2D gray‐scale image to learn effective representation of complex actions. In the later stage, we propose a voting‐based method for direction encoding to anticipate probable motion in steady images. Finally, our multiperson interaction recognition algorithm identifies which human pairs are interacting with each other using an interaction parameter. We evaluate our approach on two real‐world data sets, that is, UT‐interaction and SBU kinect interaction. The empirical experiments show that results are better than the state‐of‐the‐art methods with recognition accuracy of 95.83% on UT‐I set 1, 92.5% on UT‐I set 2, and 94.28% on SBU clean data set.

GPU-Accelerated Vision for Robots: Improving System Throughput Using OpenCV and CUDA

OpenCV is an open source computer vision and machine learning library for C/C++/Python available for Windows, Linux, macOS, and Android platforms. It contains low-level image processing functions as well as high-level algorithms such as object identification, face recognition, and action classification in videos. OpenCV has become very popular, with more than 47,000 people in its user community and 18 million downloads (see https://opencv.org/about/). Under a Berkeley Software Distribution (BSD) license, it can be used for both academic and commercial applications.

Biased Mixtures of Experts: Enabling Computer Vision Inference Under Data Transfer Limitations

We propose a novel mixture-of-experts class to optimize computer vision models in accordance with data transfer limitations at test time. Our approach postulates that the minimum acceptable amount of data allowing for highly-accurate results can vary for different input space partitions. Therefore, we consider mixtures where experts require different amounts of data, and train a sparse gating function to divide the input space for each expert. By appropriate hyperparameter selection, our approach is able to bias mixtures of experts towards selecting specific experts over others. In this way, we show that the data transfer optimization between visual sensing and processing can be solved as a convex optimization problem.To demonstrate the relation between data availability and performance, we evaluate biased mixtures on a range of mainstream computer vision problems, namely: (i) single shot detection, (ii) image super resolution, and (iii) realtime video action classification. For all cases, and when experts constitute modified baselines to meet different limits on allowed data utility, biased mixtures significantly outperform previous work optimized to meet the same constraints on available data.

Action Recognition in Still Images using Residual Neural Network Features

Action recognition in still images is central to various problems like object recognition, image retrieval, scene recognition, pose estimation and understanding of actions in videos. Human actions and behaviour in images are recognized in this problem. Images have no spatiotemporal features. So the process of action classification in videos is not applied to images. The existing works of still image action classification are reviewed. An image action recognition model is proposed in which the feature extraction is done using a deep neural network such as residual neural network and Support Vector Machine(SVM) acts as the classifier. The proposed model is evaluated with standard benchmark datasets like Pascal VOC action dataset, and Stanford actions 40 datasets using the evaluation measures such as precision, recall,f1-score, and Kappa measure. The model has high performance in Pascal VOC dataset. The model is outperformed in fifteen classes and achieved the best performance compared to other works.

Sparsity-inducing dictionaries for effective action classification

Action recognition in unconstrained videos is one of the most important challenges in computer vision. In this paper, we propose sparsity-inducing dictionaries as an effective representation for action classification in videos. We demonstrate that features obtained from sparsity based representation provide discriminative information useful for classification of action videos into various action classes. We show that the constructed dictionaries are distinct for a large number of action classes resulting in a significant improvement in classification accuracy on the HMDB51 dataset. We further demonstrate the efficacy of dictionaries and sparsity based classification on other large action video datasets like UCF50.