Action Recognition Using Deep 3D CNNs with Sequential Feature Aggregation and Attention

Action recognition has already been a heated research topic recently, which attempts to classify different human actions in videos. The current main-stream methods generally utilize ImageNet-pretrained model as features extractor, however it’s not the optimal choice to pretrain a model for classifying videos on a huge still image dataset. What’s more, very few works notice that 3D convolution neural network(3D CNN) is better for low-level spatial-temporal features extraction while recurrent neural network(RNN) is better for modelling high-level temporal feature sequences. Consequently, a novel model is proposed in our work to address the two problems mentioned above. First, we pretrain 3D CNN model on huge video action recognition dataset Kinetics to improve generality of the model. And then long short term memory(LSTM) is introduced to model the high-level temporal features produced by the Kinetics-pretrained 3D CNN model. Our experiments results show that the Kinetics-pretrained model can generally outperform ImageNet-pretrained model. And our proposed network finally achieve leading performance on UCF-101 dataset.

When using convolutional neural networks (CNNs) for segmentation of organs and lesions in medical images, the conventional approach is to work with inputs and outputs either as single slice [two-dimensional (2D)] or whole volumes [three-dimensional (3D)]. One common alternative, in this study denoted as pseudo-3D, is to use a stack of adjacent slices as input and produce a prediction for at least the central slice. This approach gives the network the possibility to capture 3D spatial information, with only a minor additional computational cost. In this study, we systematically evaluate the segmentation performance and computational costs of this pseudo-3D approach as a function of the number of input slices, and compare the results to conventional end-to-end 2D and 3D CNNs, and to triplanar orthogonal 2D CNNs. The standard pseudo-3D method regards the neighboring slices as multiple input image channels. We additionally design and evaluate a novel, simple approach where the input stack is a volumetric input that is repeatably convolved in 3D to obtain a 2D feature map. This 2D map is in turn fed into a standard 2D network. We conducted experiments using two different CNN backbone architectures and on eight diverse data sets covering different anatomical regions, imaging modalities, and segmentation tasks. We found that while both pseudo-3D methods can process a large number of slices at once and still be computationally much more efficient than fully 3D CNNs, a significant improvement over a regular 2D CNN was only observed with two of the eight data sets. triplanar networks had the poorest performance of all the evaluated models. An analysis of the structural properties of the segmentation masks revealed no relations to the segmentation performance with respect to the number of input slices. A post hoc rank sum test which combined all metrics and data sets yielded that only our newly proposed pseudo-3D method with an input size of 13 slices outperformed almost all methods. In the general case, multislice inputs appear not to improve segmentation results over using 2D or 3D CNNs. For the particular case of 13 input slices, the proposed novel pseudo-3D method does appear to have a slight advantage across all data sets compared to all other methods evaluated in this work.

Recognition of human activity based on convolutional neural network has received the interest of researchers in recent years due to its significant improvement in accuracy. A large number of algorithms based on the deep learning approach have been proposed for activity recognition purpose. However, with the increasing advancements in technologies having limited computational resources, it needs to design an efficient deep learning-based approaches with improved utilization of computational resources. This paper presents a simple and efficient 2-Dimensional convolutional neural network (2-D CNN) architecture with very small size convolutional kernel for human activity recognition. The merit of the proposed CNN architecture over standard deep learning architectures is fewer trainable parameters and lesser memory requirement which enables it to train the proposed CNN architecture on low GPU memory-based devices and also works well with smaller as well as larger size datasets. The proposed approach consists of mainly four stages: namely (1) creation of dataset and data augmentation, (2) designing 2-D convolutional neural network (CNN) architecture, (3) the proposed 2-D CNN architecture trained from scratch up to optimum stage, and (4) evaluation of the trained 2-D CNN architecture. To illustrate the effectiveness of the proposed architecture several extensive experiments are conducted on three publicly available datasets, namely IXMAS, YouTube, and UCF101 dataset. The results of the proposed method and its comparison with other state-of-the-art methods [8-12,14,18-26,29-33] demonstrate the usefulness of the proposed method.

Over large areas, land cover classification has conventionally been undertaken using satellite time series. Typically temporal metric percentiles derived from single pixel location time series have been used to take advantage of spectral differences among land cover classes over time and to minimize the impact of missing observations. Deep convolutional neural networks (CNNs) have demonstrated potential for land cover classification of single date images. However, over large areas and using time series their application is complicated because they are sensitive to missing observations and they may misclassify small and spatially fragmented surface features due to their spatial patch-based implementation. This study demonstrates, for the first time, a one-dimensional (1D) CNN single pixel time series land classification approach that uses temporal percentile metrics and that does not have these issues. This is demonstrated for all the Conterminous United States (CONUS) considering two different 1D CNN structures with 5 and 8 layers, respectively. CONUS 30 m land cover classifications were derived using all the available Landsat-5 and -7 imagery over a seven-month growing season in 2011 with 3.3 million 30 m land cover class labelled samples extracted from the contemporaneous CONUS National Land Cover Database (NLCD) 16 class land cover product. The 1D CNNs and, a conventional random forest model, were trained using 10%, 50% and 90% samples, and the classification accuracies were evaluated with an independent 10% proportion. Temporal metrics were classified using 5, 7 and 9 percentiles for each of five Landsat reflective wavelength bands and their eight band ratios. The CONUS and detailed 150 × 150 km classification results demonstrate that the approach is effective at scale and locally. The 1D CNN classification land cover class boundaries were preserved for small axis dimension features, such as roads and rivers, with no stripes or anomalous spatial patterns. The 8-layer 1D CNN provided the highest overall classification accuracies and both the 5-layer and 8-layer 1D CNN architectures provided higher accuracies than the random forest by 1.9% - 2.8% which as all the accuracies were > 83% is a meaningful increase. The CONUS overall classification accuracies increased marginally with the number of percentiles (86.21%, 86.40%, and 86.43% for 5, 7 and 9 percentiles, respectively) using the 8-layer 1D-CNN. Class specific producer and user accuracies were quantified, with lower accuracies for the developed land, crop and pasture/hay classes, but no systematic pattern among classes with respect to the number of temporal percentiles used. Application of the trained model to a different year of CONUS Landsat ARD showed moderately decreased accuracy (80.79% for 7 percentiles) that we illustrate is likely due to different intra-annual surface variations between years. These encouraging results are discussed with recommended research for deep learning using temporal metric percentiles.

Recently, three dimensional (3D) convolutional neural networks (CNNs) have emerged as dominant methods to capture spatiotemporal representations in videos, by adding to pre-existing 2D CNNs a third, temporal dimension. Such 3D CNNs, however, are anti-causal (i.e., they exploit information from both the past and the future frames to produce feature representations, thus preventing their use in online settings), constrain the temporal reasoning horizon to the size of the temporal convolution kernel, and are not temporal resolution-preserving for video sequence-to-sequence modelling, as, for instance, in action detection. To address these serious limitations, here we present a new 3D CNN architecture for the causal/online processing of videos. Namely, we propose a novel Recurrent Convolutional Network (RCN), which relies on recurrence to capture the temporal context across frames at each network level. Our network decomposes 3D convolutions into (1) a 2D spatial convolution component, and (2) an additional hidden state 1 × 1 convolution, applied across time. The hidden state at any time t is assumed to depend on the hidden state at t - 1 and on the current output of the spatial convolution component. As a result, the proposed network: (i) produces causal outputs, (ii) provides flexible temporal reasoning, (iii) preserves temporal resolution. Our experiments on the largescale large Kinetics and MultiThumos datasets show that the proposed method performs comparably to anti-causal 3D CNNs, while being causal and using fewer parameters.

The past several years have seen both an explosion in the use of Convolutional Neural Networks (CNNs) and accelerators to make CNN inference practical. In the architecture community, the lion share of effort has targeted CNN inference for image recognition. The closely related problem of video recognition has received far less attention as an accelerator target. This is surprising, as video recognition is more computationally intensive than image recognition, and video traffic is predicted to be the majority of internet traffic in the coming years. This paper fills the gap between algorithmic and hardware advances for video recognition by providing a design space exploration and flexible architecture for accelerating 3D Convolutional Neural Networks (3D CNNs)—the core kernel in modern video understanding. When compared to (2D) CNNs used for image recognition, efficiently accelerating 3D CNNs poses a significant engineering challenge due to their large (and variable over time) memory footprint and higher dimensionality. To address these challenges, we design a novel accelerator called Morph, that can adaptively support different spatial and temporal tiling strategies depending on the needs of each layer of each target 3D CNN. We codesign a software infrastructure alongside the Morph hardware to find good-fit parameters to control the hardware. Evaluated on state-of-the-art 3D CNNs, Morph achieves up to 2.7× (1.9× average) reduction in energy consumption and improves performance/watt up to 4.4× (3× average) compared to a baseline 3D CNN accelerator, with an area overhead of 2%. Morph further achieves a 11.6× average energy reduction on 3D CNNs when compared to Eyeriss, a popular 2D CNN accelerator, while reducing efficiency compared to Eyeriss on a 2D CNN by 71%.

Action recognition is to automatically detect and classify human’s action in videos, with difficulty lies in modeling temporal relationship between frame sequences. The well-used 2D convolution neural network (CNN) is not suitable for this work, due to lacking temporal modeling ability. In this paper, a novel 2D CNN with inter frame information extraction module based on bilinear operation is proposed to deal with this problem. This model can greatly improve the temporal modeling ability of 2D CNN and just introduce a small amount of storage and calculation via parameter decomposition method. In addition, it has a flexible form to easily make tradeoff between performance and complexity. Finally, the effectiveness of this new network is validated on two kinds of benchmarks including both temporal-related (Something-Something v1) and scene-related(mini-kinetics), with top-1 accuracy 44.5% and 67.8% respectively, which reach or exceed the performance of existing methods with the similar model complexity.

Recently, several convolutional neural networks have been proposed not only for 2D images, but also for 3D and 4D volume segmentation. Nevertheless, due to the large data size of the latter, acquiring a sufficient amount of training annotations is much more strenuous than in 2D images. For 4D time-series tomograms, this is usually handled by segmenting the constituent tomograms independently through time with 3D convolutional neural networks. Inter-volume information is therefore not utilized, potentially leading to temporal incoherence. In this paper, we attempt to resolve this by proposing two hidden Markov model variants that refine 4D segmentation labels made by 3D convolutional neural networks working on each time point. Our models utilize not only inter-volume information, but also the prediction confidence generated by the 3D segmentation convolutional neural networks themselves. To the best of our knowledge, this is the first attempt to refine 4D segmentations made by 3D convolutional neural networks using hidden Markov models. During experiments we test our models, qualitatively, quantitatively and behaviourally, using prespecified segmentations. We demonstrate in the domain of time series tomograms which are typically undersampled to allow more frequent capture; a particularly challenging problem. Finally, our dataset and code is publicly available.

The convolutional neural network (CNN) is a potent and popular neural network types and has been crucial to deep learning in recent years. A standard CNN which is known as 2-dimensions CNN was first proposed to solve image classification problems at first. With advancements in science and technology, as well as the growth of the internet, tasks involving video analysis have received increased attention in recent years. As a result, it was suggested that a 3-dimensions convolutional neural network (3D CNN) be applied to tackle 3D image processing or video analysis. 3D CNN is primarily used in the field of human activity recognition, video classification, medical imaging detection, and so on. This paper offers a succinct introduction to the general architecture and representative models of 3D CNN, contrasts the variations between 2D CNN and 3D CNN, explores the widely used models derived from 3D CNN, and displays the application outcomes of 3D CNN.

Mobile devices are becoming an important carrier for deep learning tasks, as they are being equipped with powerful, high-end mobile CPUs and GPUs. However, it is still a challenging task to execute 3D Convolutional Neural Networks (CNNs) targeting for real-time performance, besides high inference accuracy. The reason is more complex model structure and higher model dimensionality overwhelm the available computation/storage resources on mobile devices. A natural way may be turning to deep learning weight pruning techniques. However, the direct generalization of existing 2D CNN weight pruning methods to 3D CNNs is not ideal for fully exploiting mobile parallelism while achieving high inference accuracy. This paper proposes RT3D, a model compression and mobile acceleration framework for 3D CNNs, seamlessly integrating neural network weight pruning and compiler code generation techniques. We propose and investigate two structured sparsity schemes i.e., the vanilla structured sparsity and kernel group structured (KGS) sparsity that are mobile acceleration friendly. The vanilla sparsity removes whole kernel groups, while KGS sparsity is a more fine-grained structured sparsity that enjoys higher flexibility while exploiting full on-device parallelism. We propose a reweighted regularization pruning algorithm to achieve the proposed sparsity schemes. The inference time speedup due to sparsity is approaching the pruning rate of the whole model FLOPs (floating point operations). RT3D demonstrates up to 29.1x speedup in end-to-end inference time comparing with current mobile frameworks supporting 3D CNNs, with moderate 1%~1.5% accuracy loss. The end-to-end inference time for 16 video frames could be within 150 ms, when executing representative C3D and R(2+1)D models on a cellphone. For the first time, real-time execution of 3D CNNs is achieved on off-the-shelf mobiles.

Pre-trained vision-language models, exemplified by CLIP, have exhibited promising zero-shot capabilities across various downstream tasks. Trained on image-text pairs, CLIP is naturally extendable to video-based action recognition, due to the similarity between processing images and video frames. To leverage this inherent synergy, numerous efforts have been directed towards adapting CLIP for action recognition tasks in videos. However, the specific methodologies for this adaptation remain an open question. Common approaches include prompt tuning and fine-tuning with or without extra model components on video-based action recognition tasks. Nonetheless, such adaptations may compromise the generalizability of the original CLIP framework and also necessitate the acquisition of new training data, thereby undermining its inherent zero-shot capabilities. In this study, we propose zero-shot action recognition (ZAR) by adapting the CLIP pre-trained model without the need for additional training datasets. Our approach leverages the entropy minimization technique, utilizing the current video test sample and augmenting it with varying frame rates. We encourage the model to make consistent decisions, and use this consistency to dynamically update a prompt learner during inference. Experimental results demonstrate that our ZAR method achieves state-of-the-art zero-shot performance on the Kinetics-600, HMDB51, and UCF101 datasets.

Multimodal emotion recognition has become a hot topic in human-computer interaction and intelligent healthcare fields. However, combining information from different human different modalities for emotion computation is still challenging. In this paper, we propose a three-dimensional convolutional recurrent neural network model (referred to as 3FACRNN network) based on multimodal fusion and attention mechanism. The 3FACRNN network model consists of a visual network and an EEG network. The visual network is composed of a cascaded convolutional neural network-time convolutional network (CNN-TCN). In the EEG network, the 3D feature building module was added to integrate band information, spatial information and temporal information of the EEG signal, and the band attention and self-attention modules were added to the convolutional recurrent neural network (CRNN). The former explores the effect of different frequency bands on network recognition performance, while the latter is to obtain the intrinsic similarity of different EEG samples. To investigate the effect of different frequency bands on the experiment, we obtained the average attention mask for all subjects in different frequency bands. The distribution of the attention masks across the different frequency bands suggests that signals more relevant to human emotions may be active in the high frequency bands γ (31-50 Hz). Finally, we try to use the multi-task loss function Lc to force the approximation of the intermediate feature vectors of the visual and EEG modalities, with the aim of using the knowledge of the visual modalities to improve the performance of the EEG network model. The mean recognition accuracy and standard deviation of the proposed method on the two multimodal sentiment datasets DEAP and MAHNOB-HCI (arousal, valence) were 96.75 ± 1.75, 96.86 ± 1.33; 97.55 ± 1.51, 98.37 ± 1.07, better than those of the state-of-the-art multimodal recognition approaches. The experimental results show that starting from the multimodal information, the facial video frames and electroencephalogram (EEG) signals of the subjects are used as inputs to the emotion recognition network, which can enhance the stability of the emotion network and improve the recognition accuracy of the emotion network. In addition, in future work, we will try to utilize sparse matrix methods and deep convolutional networks to improve the performance of multimodal emotion networks.

An action recognition network that combines multi-level spatiotemporal feature fusion with an attention mechanism is proposed as a solution to the issues of single spatiotemporal feature scale extraction, information redundancy, and insufficient extraction of frequency domain information in channels in 3D convolutional neural networks. Firstly, based on 3D CNN, this paper designs a new multilevel spatiotemporal feature fusion (MSF) structure, which is embedded in the network model, mainly through multilevel spatiotemporal feature separation, splicing and fusion, to achieve the fusion of spatial perceptual fields and short-medium-long time series information at different scales with reduced network parameters; In the second step, a multi-frequency channel and spatiotemporal attention module (FSAM) is introduced to assign different frequency features and spatiotemporal features in the channels are assigned corresponding weights to reduce the information redundancy of the feature maps. Finally, we embed the proposed method into the R3D model, which replaced the 2D convolutional filters in the 2D Resnet with 3D convolutional filters and conduct extensive experimental validation on the small and medium-sized dataset UCF101 and the large-sized dataset Kinetics-400. The findings revealed that our model increased the recognition accuracy on both datasets. Results on the UCF101 dataset, in particular, demonstrate that our model outperforms R3D in terms of a maximum recognition accuracy improvement of 7.2% while using 34.2% fewer parameters. The MSF and FSAM are migrated to another traditional 3D action recognition model named C3D for application testing. The test results based on UCF101 show that the recognition accuracy is improved by 8.9%, proving the strong generalization ability and universality of the method in this paper.

Idiopathic Parkinsons disease and atypical parkinsonian syndromes have similar symptoms at early disease stages, which makes the early differential diagnosis difficult. Positron emission tomography with 18F-FDG shows the ability to assess early neuronal dysfunction of neurodegenerative diseases and is well established for clinical use. In the past decades, machine learning methods have been widely used for the differential diagnosis of parkinsonism based on metabolic patterns. Unlike these conventional machine learning methods relying on hand-crafted features, the deep convolutional neural networks, which have achieved significant success in medical applications recently, have the advantage of learning salient feature representations automatically and effectively. This advantage may offer more appropriate invisible features extracted from data for the enhancement of the diagnosis accuracy. Therefore, this paper develops a 3D deep convolutional neural network on 18F-FDG PET images for the automated early diagnosis. Furthermore, we depicted in saliency maps the decision mechanism of the deep learning method to assist the physiological interpretation of deep learning performance. The proposed method was evaluated on a dataset with 920 patients. In addition to improving the accuracy in the differential diagnosis of parkinsonism compared to state-of-the-art approaches, the deep learning methods also discovered saliency features in a number of critical regions (e.g., midbrain), which are widely accepted as characteristic pathological regions for movement disorders but were ignored in the conventional analysis of FDG PET images.

Visual attention mechanism has been widely used in computer vision and plays a vital role in the research of human action recognition. In this paper, we explore a novel moving target detection mechanism for human action recognition and propose a new 3D CNN (3D Convolutional Neural Network) model, dubbed LaM-2SRN (Local Features Enhanced and Moving target detected 2Stream-ResNet) for extracting and learning attention-enhanced spatiotemporal features. The contributions of this paper are as follow: First, the traditional CAM (Class Activation Map) based visual attention algorithm is used to obtain the optical flow information of the human region, thus eliminating the influence of irrelevant optical flow information (such as background clutter). Second, the ViBe algorithm is used to identify the moving target in the continuous frame, retain the optical flow information of the moving target, and make it complement with the optical flow information of the human region, to obtain a complete motion descriptor. After the motion information is marked in the video frame, we put the marked video frames and the original video frames into the 3D CNN and 2D CNN models respectively, to acquired mixed descriptor, which serves as the basis for video classification. Different from most of the previous target detection algorithms, this paper not only detects the human target, but also detects the moving foreground targets in the video, so that the pre-trained CNN model can obtain more complete motion information from videos. We use RGB-only video data for evaluation on two benchmarks: UCF101 and HMDB51, the experimental results demonstrate that our LaM-2SRN is comparable to the previous state-of-the-art algorithms.