Long-range Information Research Articles

GTMFuse: Group-attention transformer-driven multiscale dense feature-enhanced network for infrared and visible image fusion

Infrared and visible images captured by different devices can be seamlessly integrated into a single composite image through the application of image fusion techniques. However, many existing convolutional neural network-based methods for infrared and visible image fusion have exhibited limited capability for effectively amalgamating information from the source images. Consequently, we propose a group-attention transformer into the multiscale feature enhanced network for infrared and visible image fusion, which we abbreviate as GTMFuse. Specifically, GTMFuse employs multiscale dual-channel encoders to independently process the source image and extract multiscale features. Among the encoders, the group-attention transformer module is utilized to facilitate more comprehensive long-range feature dependency modeling at each scale. This innovative module seamlessly combines a fixed-direction stripe attention mechanism with channel attention and window attention, enabling comprehensive global long-range information capture and interaction with feature information across the source images. The multiscale features obtained from the group-attention transformer module are integrated into the fused image through a meticulously designed dense fusion block. Furthermore, this study introduces a novel dataset named HBUT-IV, encompassing surveillance images captured from multiple viewpoints. The HBUT-IV dataset serves as a valuable benchmark for assessing the efficacy of fusion methods. Extensive experiments are conducted on four datasets employing nine comparative methods, revealing the superior performance of the GTMFuse approach. The implementation code is accessible at https://github.com/XingLongH/GTMFuse.

Usage of an improved YOLOv5 for steel surface defect detection

Abstract The one-stage YOLOv5 steel surface defect detection has issues such as slow operation speed, loss of defect location and semantic information of small targets, and inadequate extraction of defect features. This study proposed a defect detection algorithm with improved YOLOv5 to solve these issues. The proposed algorithm used the slim-neck layer built by three new modules instead of the neck layer in YOLOv5s to achieve a lightweight network model. In addition, the spatial perception self-attention mechanism was introduced to enhance the feature extraction capability of the initial convolutional layer without limiting the input size. The improved Atrous Spatial Pyramid Pooling was added to expand the perceptual field and capture multiscale contextual information while preventing local information loss and enhancing the relevance of long-range information. The experimental results showed that the improved YOLOv5 algorithm has a reduced model volume, significantly higher detection accuracy and speed than the traditional algorithm, and the ability to detect steel surface defects quickly and accurately.

Shots segmentation-based optimized dual-stream framework for robust human activity recognition in surveillance video

Nowadays, for controlling crime, surveillance cameras are typically installed in all public places to ensure urban safety and security. However, automating Human Activity Recognition (HAR) using computer vision techniques faces several challenges such as lowlighting, complex spatiotemporal features, clutter backgrounds, and inefficient utilization of surveillance system resources. Existing attempts in HAR designed straightforward networks by analyzing either spatial or motion patterns resulting in limited performance while the dual streams methods are entirely based on Convolutional Neural Networks (CNN) that are inadequate to learning the long-range temporal information for HAR. To overcome the above-mentioned challenges, this paper proposes an optimized dual stream framework for HAR which mainly consists of three steps. First, a shots segmentation module is introduced in the proposed framework to efficiently utilize the surveillance system resources by enhancing the lowlight video stream and then it detects salient video frames that consist of human. This module is trained on our own challenging Lowlight Human Surveillance Dataset (LHSD) which consists of both normal and different levels of lowlighting data to recognize humans in complex uncertain environments. Next, to learn HAR from both contextual and motion information, a dual stream approach is used in the feature extraction. In the first stream, it freezes the learned weights of the backbone Vision Transformer (ViT) B-16 model to select the discriminative contextual information. In the second stream, ViT features are then fused with the intermediate encoder layers of FlowNet2 model for optical flow to extract a robust motion feature vector. Finally, a two stream Parallel Bidirectional Long Short-Term Memory (PBiLSTM) is proposed for sequence learning to capture the global semantics of activities, followed by Dual Stream Multi-Head Attention (DSMHA) with a late fusion strategy to optimize the huge features vector for accurate HAR. To assess the strength of the proposed framework, extensive empirical results are conducted on real-world surveillance scenarios and various benchmark HAR datasets that achieve 78.6285%, 96.0151%, and 98.875% accuracies on HMDB51, UCF101, and YouTube Action, respectively. Our results show that the proposed strategy outperforms State-of-the-Art (SOTA) methods. The proposed framework gives superior performance in HAR, providing accurate and reliable recognition of human activities in surveillance systems.

Dynamic correlations: exact and approximate methods for mutual information.

Proteins are dynamic entities that undergo conformational changes critical for their functions. Understanding the communication pathways and information transfer within proteins is crucial for elucidating allosteric interactions in their mechanisms. This study utilizes mutual information (MI) analysis to probe dynamic allostery. Using two cases, Ubiquitin and PLpro, we have evaluated the accuracy and limitations of different approximations including the exact anisotropic and isotropic models, multivariate Gaussian model, isotropic Gaussian model, and the Gaussian Network Model (GNM) in revealing allosteric interactions. Our findings emphasize the required trajectory length for capturing accurate mutual information profiles. Long molecular dynamics trajectories, 1 ms for Ubiquitin and 100 µs for PLpro are used as benchmarks, assuming they represent the ground truth. Trajectory lengths of approximately 5 µs for Ubiquitin and 1 µs for PLpro marked the onset of convergence, while the multivariate Gaussian model accurately captured mutual information with trajectories of 5 ns for Ubiquitin and 350 ns for PLpro. However, the isotropic Gaussian model is less successful in representing the anisotropic nature of protein dynamics, particularly in the case of PLpro, highlighting its limitations. The GNM, however, provides reasonable approximations of long-range information exchange as a minimalist network model based on a single crystal structure. Overall, the optimum trajectory lengths for effective Gaussian approximations of long-time dynamic behavior depend on the inherent dynamics within the protein's topology. The GNM, by showcasing dynamics across relatively diverse time scales, can be used either as a standalone method or to gauge the adequacy of MD simulation lengths. Mutual information codes are available at https://github.com/kemaldemirtas/prc-MI.git.

NAGNE: Node-to-Attribute Generation Network Embedding for Heterogeneous Network

Heterogeneous network embedding aims to project multiple types of nodes into a low-dimensional space, and has become increasingly ubiquitous. However, several challenges have not been addressed so far. First, existing heterogeneous network embedding techniques typically rely on meta-paths to deal with the complex heterogeneous network. Using these meta-paths requires prior knowledge from domain experts for optimal meta-path selection. Second, few existing models can effectively consider both heterogeneous structural information and heterogeneous node attribute information. Third, existing models preserve the structure information by considering the first- and/or the second-order proximities, which cannot capture long-range structural information. To address these limitations, we propose a novel attributed heterogeneous network embedding model referred to as Node-to-Attribute Generation Network Embedding (NAGNE). NAGNE comprises two major components, the attributed random walk which samples node sequences in attributed heterogeneous network, and the node-to-attribute generation which learns the mapping that translates each node sequence itself from the node sequence to the node attribute sequence. Extensive experiments on three heterogeneous network datasets demonstrate that NAGNE outperforms state-of-the-art baselines in various data mining tasks.

Attenphos: General Phosphorylation Site Prediction Model Based on Attention Mechanism.

Phosphorylation site prediction has important application value in the field of bioinformatics. It can act as an important reference and help with protein function research, protein structure research, and drug discovery. So, it is of great significance to propose scientific and effective calculation methods to accurately predict phosphorylation sites. In this study, we propose a new method, Attenphos, based on the self-attention mechanism for predicting general phosphorylation sites in proteins. The method not only captures the long-range dependence information of proteins but also better represents the correlation between amino acids through feature vector encoding transformation. Attenphos takes advantage of the one-dimensional convolutional layer to reduce the number of model parameters, improve model efficiency and prediction accuracy, and enhance model generalization. Comparisons between our method and existing state-of-the-art prediction tools were made using balanced datasets from human proteins and unbalanced datasets from mouse proteins. We performed prediction comparisons using independent test sets. The results showed that Attenphos demonstrated the best overall performance in the prediction of Serine (S), Threonine (T), and Tyrosine (Y) sites on both balanced and unbalanced datasets. Compared to current state-of-the-art methods, Attenphos has significantly higher prediction accuracy. This proves the potential of Attenphos in accelerating the identification and functional analysis of protein phosphorylation sites and provides new tools and ideas for biological research and drug discovery.

Single-residue linear and conformational B cell epitopes prediction using random and ESM-2 based projections.

B cell epitope prediction methods are separated into linear sequence-based predictors and conformational epitope predictions that typically use the measured or predicted protein structure. Most linear predictions rely on the translation of the sequence to biologically based representations and the applications of machine learning on these representations. We here present CALIBER 'Conformational And LInear B cell Epitopes pRediction', and show that a bidirectional long short-term memory with random projection produces a more accurate prediction (test set AUC=0.789) than all current linear methods. The same predictor when combined with an Evolutionary Scale Modeling-2 projection also improves on the state of the art in conformational epitopes (AUC = 0.776). The inclusion of the graph of the 3D distances between residues did not increase the prediction accuracy. However, the long-range sequence information was essential for high accuracy. While the same model structure was applicable for linear and conformational epitopes, separate training was required for each. Combining the two slightly increased the linear accuracy (AUC 0.775 versus 0.768) and reduced the conformational accuracy (AUC = 0.769).

Multi-phase CT synthesis-assisted segmentation of abdominal organs

To propose a method for abdominal multi-organ segmentation assisted by multi-phase CT synthesis. Multi-phase CT synthesis for synthesizing high-quality CT images was used to increase the information details for image segmentation. A transformer block was introduced to help to capture long-range semantic information in cooperation with perceptual loss to minimize the differences between the real image and synthesized image. The model was trained using multi-phase CT dataset of 526 total cases from Nanfang Hospital. The mean maximum absolute error (MAE) of the synthesized non-contrast CT, venous phase contrast- enhanced CT (CECT), and delay phase CECT images from arterial phase CECT was 19.192±3.381, 20.140±2.676 and 22.538±2.874, respectively, which were better than those of images synthesized using other methods. Validation of the multi-phase CT synthesis-assisted abdominal multi-organ segmentation method showed an average dice coefficient of 0.847 for the internal validation set and 0.823 for the external validation set. The propose method is capable of synthesizing high-quality multi-phase CT images to effectively reduce the errors in registration between different phase CT images and improve the performance for segmentation of 13 abdominal organs.

The series of conformational states adopted by rotorless F1-ATPase during its hydrolysis cycle

F1Fo ATP synthase interchanges phosphate transfer energy and proton motive force via a rotary catalytic mechanism and isolated F1-ATPase subcomplexes can also hydrolyze ATP to generate rotation of their central γ rotor subunit. As ATP is hydrolyzed, the F1-ATPase cycles through a series of conformational states that mediates unidirectional rotation of the rotor. However, even in the absence of a rotor, the α and β subunits are still able to pass through a series of conformations, akin to those that generate rotation. Here, we use cryoelectron microscopy to establish the structures of these rotorless states. These structures indicate that cooperativity in this system is likely mediated by contacts between the β subunit lever domains, irrespective of the presence of the γ rotor subunit. These findings provide insight into how long-range information may be transferred in large biological systems.

A novel transformer autoencoder for multi-modal emotion recognition with incomplete data

Multi-modal signals have become essential data for emotion recognition since they can represent emotions more comprehensively. However, in real-world environments, it is often impossible to acquire complete data on multi-modal signals, and the problem of missing modalities causes severe performance degradation in emotion recognition. Therefore, this paper represents the first attempt to use a transformer-based architecture, aiming to fill the modality-incomplete data from partially observed data for multi-modal emotion recognition (MER). Concretely, this paper proposes a novel unified model called transformer autoencoder (TAE), comprising a modality-specific hybrid transformer encoder, an inter-modality transformer encoder, and a convolutional decoder. The modality-specific hybrid transformer encoder bridges a convolutional encoder and a transformer encoder, allowing the encoder to learn local and global context information within each particular modality. The inter-modality transformer encoder builds and aligns global cross-modal correlations and models long-range contextual information with different modalities. The convolutional decoder decodes the encoding features to produce more precise recognition. Besides, a regularization term is introduced into the convolutional decoder to force the decoder to fully leverage the complete and incomplete data for emotional recognition of missing data. 96.33%, 95.64%, and 92.69% accuracies are attained on the available data of the DEAP and SEED-IV datasets, and 93.25%, 92.23%, and 81.76% accuracies are obtained on the missing data. Particularly, the model acquires a 5.61% advantage with 70% missing data, demonstrating that the model outperforms some state-of-the-art approaches in incomplete multi-modal learning.

Enhancing artistic analysis through deep learning: a graphic art element recognition model based on SSD and FPT.

For the analysis of art works, accurate identification of various elements of works through deep learning methods is helpful for artists to appreciate and learn works. In this study, we leverage deep learning methodologies to precisely identify the diverse elements within graphic art designs, aiding artists in their appreciation and learning process. Our approach involves integrating the attention mechanism into an enhanced Single Shot MultiBox Detector (SSD) model to refine the recognition of artistic design elements. Additionally, we improve the feature fusion structure of the SSD model by incorporating long-range attention mechanism information, thus enhancing target detection accuracy. Moreover, we refine the Feature Pyramid Transformer (FPT) attention mechanism model to ensure the output feature map aligns effectively with the requirements of object detection. Our empirical findings demonstrate that our refined approach outperforms the original SSD algorithm across all four evaluation metrics, exhibiting improvements of 1.52%, 1.89%, 3.09%, and 2.57%, respectively. Qualitative tests further illustrate the accuracy, robustness, and universality of our proposed method, particularly in scenarios characterized by dense artistic elements and challenging-to-distinguish categories within art compositions.

Lightweight and fast visual detection method for 3C assembly

In the context of 3C assembly scenarios, characterized by numerous semi-flexible, heterogeneous, and small slender targets, traditional target detection algorithms face significant challenges such as low accuracy, weak generalization, large model sizes, and slow inference speeds. To address these issues, this study introduces an enhanced method based on the YOLOv5 model, named YOLOv5-GTB. This method integrates Bidirectional Feature Pyramid Network (BiFPN), Ghost lightweight convolution, Vision Transformer (ViT), and adaptive activation function technologies, aiming to improve the accuracy and speed of target detection. Our approach utilizes Ghost convolution to construct a Ghost bottleneck layer, optimizing the feature extraction network while significantly reducing computational costs and enhancing the convolutional neural network’s feature extraction capabilities. Additionally, the Cross-Stage Partial Network (CSPNet) architecture is employed to effectively segment the data flow of the input feature map layer, thereby improving the efficiency of gradient processing. Furthermore, we introduce a fusion structure of CNN and Transformer, leveraging the strengths of convolutional neural networks in local feature extraction and the ViT’s capability in long-range information capture, thereby further enhancing the overall network’s feature extraction performance. Regarding feature fusion, considering the limitations of traditional top-down unidirectional information flow in effectively merging features with both location and semantic information, BiFPN is incorporated into the YOLOv5-GTB. This enhances the fusion of feature layers extracted from both the end of the backbone network and the first module’s fused feature layer, thus improving detection accuracy. Ablation and comparative experiments conducted on the 3C assembly scenario target detection dataset demonstrate the significant advantages of the YOLOv5-GTB model in terms of accuracy and speed. Ultimately, the application of this model to the 3C assembly platform successfully achieves rapid and accurate target recognition in this scenario.

SSRL: Self-Supervised Spatial-Temporal Representation Learning for 3D Action Recognition

For 3D action recognition, the main challenge is to extract long-range semantic information in both temporal and spatial dimensions. In this paper, in order to better excavate long-range semantic information from large number of unlabelled skeleton sequences, we propose Self-supervised Spatial-temporal Representation Learning (SSRL), a contrastive learning framework to learn skeleton representation. SSRL consists of two novel inference tasks that enable the network to learn global semantic information in the temporal and spatial dimensions, respectively. The temporal inference task learns the temporal persistence of human actions through temporally incomplete skeleton sequences. And the spatial inference task learns the spatially coordinated nature of human action through spatially partially skeleton sequence. We design two transformation modules to efficiently realize these two tasks while fitting the encoder network. To avoid the difficulty of constructing and maintaining high-quality negative samples, our proposed framework learns by maintaining consistency among positive samples without the need of any negative sample. Experiments demonstrate that our proposed method can achieve better results in comparison with state-of-the-art methods under a variety of evaluation protocols on NTU RGB+D 60, PKU-MMD and NTU RGB+D 120 datasets.

An Efficient Multiscale Spatial Rearrangement MLP Architecture for Image Restoration.

The effective use of long-range information can yield improved network performance, which is very important for image restoration. Although local window-based models have linear complexity and can be feasibly applied to process high-resolution images, a single-scale window has a limited receptive field and is less efficient for encoding long-range context information. To address this issue, this paper presents a single-stage multiscale spatial rearrangement multilayer perceptron (MSSR-MLP) architecture that can obtain information at different scales within a local window. Specifically, we propose a simple and efficient spatial rearrangement module (SRM) that moves information outside the local window to the inside of the local window so that long-range dependencies can be modeled using only a window-based fully connected (FC) layer. The SRM can extend the local receptive field of a window-based FC layer without introducing additional parameters and FLOPs. Utilizing several spatial rearrangement modules with different step sizes, we design an efficient multiscale spatial rearrangement MLP architecture for image restoration. This design aggregates multiscale information to achieve improved restoration quality while maintaining a low computational cost. Extensive experiments conducted on several image restoration tasks demonstrate the efficiency and effectiveness of our method. For example, it requires only ~4.3% of the FLOPs needed by SwinIR for Gaussian gray image denoising, ~13.9% of the FLOPs needed by C2 PNet for single-image dehazing and ~18.9% of the FLOPs needed by MAXIM for single-image motion deblurring but achieves better performance on each of these restoration tasks.

Spatial Prior-Guided Bi-Directional Cross-Attention Transformers for Tooth Instance Segmentation.

Tooth instance segmentation of dental panoramic X-ray images represents a task of significant clinical importance. Teeth demonstrate symmetry within the upper and lower jawbones and are arranged in a specific order. However, previous studies frequently overlook this crucial spatial prior information, resulting in misidentifications of tooth categories for adjacent or similarly shaped teeth. In this paper, we propose SPGTNet, a spatial prior-guided transformer method, designed to both the extracted tooth positional features from CNNs and the long-range contextual information from vision transformers for dental panoramic X-ray image segmentation. Initially, a center-based spatial prior perception module is employed to identify each tooth's centroid, thereby enhancing the spatial prior information for the CNN sequence features. Subsequently, a bi-directional cross-attention module is designed to facilitate the interaction between the spatial prior information of the CNN sequence features and the long-distance contextual features of the vision transformer sequence features. Finally, an instance identification head is employed to derive the tooth segmentation results. Extensive experiments on three public benchmark datasets have demonstrated the effectiveness and superiority of our proposed method in comparison with other state-of-the-art approaches. The proposed method demonstrates the capability to accurately identify and analyze tooth structures, thereby providing crucial information for dental diagnosis, treatment planning, and research.

Learning Motion and Temporal Cues for Unsupervised Video Object Segmentation.

In this article, we address the challenges in unsupervised video object segmentation (UVOS) by proposing an efficient algorithm, termed MTNet, which concurrently exploits motion and temporal cues. Unlike previous methods that focus solely on integrating appearance with motion or on modeling temporal relations, our method combines both aspects by integrating them within a unified framework. MTNet is devised by effectively merging appearance and motion features during the feature extraction process within encoders, promoting a more complementary representation. To capture the intricate long-range contextual dynamics and information embedded within videos, a temporal transformer module is introduced, facilitating efficacious interframe interactions throughout a video clip. Furthermore, we employ a cascade of decoders all feature levels across all feature levels to optimally exploit the derived features, aiming to generate increasingly precise segmentation masks. As a result, MTNet provides a strong and compact framework that explores both temporal and cross-modality knowledge to robustly localize and track the primary object accurately in various challenging scenarios efficiently. Extensive experiments across diverse benchmarks conclusively show that our method not only attains state-of-the-art performance in UVOS but also delivers competitive results in video salient object detection (VSOD). These findings highlight the method's robust versatility and its adeptness in adapting to a range of segmentation tasks. The source code is available at https://github.com/hy0523/MTNet.

GA2MIF: Graph and Attention Based Two-Stage Multi-Source Information Fusion for Conversational Emotion Detection

Multimodal Emotion Recognition in Conversation (ERC) plays an influential role in the field of human-computer interaction and conversational robotics since it can motivate machines to provide empathetic services. Multimodal data modeling is an up-and-coming research area in recent years, which is inspired by human capability to integrate multiple senses. Several graph-based approaches claim to capture interactive information between modalities, but the heterogeneity of multimodal data makes these methods prohibit optimal solutions. In this work, we introduce a multimodal fusion approach named Graph and Attention based Two-stage Multi-source Information Fusion (GA2MIF) for emotion detection in conversation. Our proposed method circumvents the problem of taking heterogeneous graph as input to the model while eliminating complex redundant connections in the construction of graph. GA2MIF focuses on contextual modeling and cross-modal modeling through leveraging Multi-head Directed Graph ATtention networks (MDGATs) and Multi-head Pairwise Cross-modal ATtention networks (MPCATs), respectively. Extensive experiments on two public datasets (i.e., IEMOCAP and MELD) demonstrate that the proposed GA2MIF has the capacity to validly capture intra-modal long-range contextual information and inter-modal complementary information, as well as outperforms the prevalent State-Of-The-Art (SOTA) models by a remarkable margin.

COSTA: A Multi-center TOF-MRA Dataset and A Style Self-Consistency Network for Cerebrovascular Segmentation.

Time-of-flight magnetic resonance angiography (TOF-MRA) is the least invasive and ionizing radiation-free approach for cerebrovascular imaging, but variations in imaging artifacts across different clinical centers and imaging vendors result in inter-site and inter-vendor heterogeneity, making its accurate and robust cerebrovascular segmentation challenging. Moreover, the limited availability and quality of annotated data pose further challenges for segmentation methods to generalize well to unseen datasets. In this paper, we construct the largest and most diverse TOF-MRA dataset (COSTA) from 8 individual imaging centers, with all the volumes manually annotated. Then we propose a novel network for cerebrovascular segmentation, namely CESAR, with the ability to tackle feature granularity and image style heterogeneity issues. Specifically, a coarse-to-fine architecture is implemented to refine cerebrovascular segmentation in an iterative manner. An automatic feature selection module is proposed to selectively fuse global long-range dependencies and local contextual information of cerebrovascular structures. A style self-consistency loss is then introduced to explicitly align diverse styles of TOF-MRA images to a standardized one. Extensive experimental results on the COSTA dataset demonstrate the effectiveness of our CESAR network against state-of-the-art methods. We have made 6 subsets of COSTA with the source code online available, in order to promote relevant research in the community.

Attention-Aware Non-Rigid Image Registration for Accelerated MR Imaging.

Accurate motion estimation at high acceleration factors enables rapid motion-compensated reconstruction in Magnetic Resonance Imaging (MRI) without compromising the diagnostic image quality. In this work, we introduce an attention-aware deep learning-based framework that can perform non-rigid pairwise registration for fully sampled and accelerated MRI. We extract local visual representations to build similarity maps between the registered image pairs at multiple resolution levels and additionally leverage long-range contextual information using a transformer-based module to alleviate ambiguities in the presence of artifacts caused by undersampling. We combine local and global dependencies to perform simultaneous coarse and fine motion estimation. The proposed method was evaluated on in-house acquired fully sampled and accelerated data of 101 patients and 62 healthy subjects undergoing cardiac and thoracic MRI. The impact of motion estimation accuracy on the downstream task of motion-compensated reconstruction was analyzed. We demonstrate that our model derives reliable and consistent motion fields across different sampling trajectories (Cartesian and radial) and acceleration factors of up to 16x for cardiac motion and 30x for respiratory motion and achieves superior image quality in motion-compensated reconstruction qualitatively and quantitatively compared to conventional and recent deep learning-based approaches. The code is publicly available at https://github.com/lab-midas/GMARAFT.

Sparse Action Tube Detection.

Action tube detection is a challenging task as it requires not only to locate action instances in each frame, but also link them in time. Existing action tube detection methods often employ multi-stage pipelines with complex designs and time-consuming linking procedure. In this paper, we present a simple end-to-end action tube detection method, termed as Sparse Tube Detector (STDet). Unlike those dense action detectors, our core idea is to use a set of learnable tube queries and directly decode them into action tubes (i.e., a set of tracked boxes with action label) from video content. This sparse detection paradigm shares several advantages. First, the large number of hand-crafted anchor candidates in dense action detectors is greatly reduced to a small number of learnable tubes, which results in a more efficient detection framework. Second, our learnable tube queries directly attend the whole video content, which endows our method with the capacity of capturing long-range information for action detection. Finally, our action detector is an end-to-end tube detection without requiring the linking procedure, which directly and explicitly predicts the action boundary instead of depending on the linking strategy. Extensive experiments shows that our STDet outperforms the previous state-of-the-art methods on two challenging untrimmed video action detection datasets of UCF101-24 and MultiSports. We hope our method will be an simple end-to-end tube detection baseline and can inspire new ideas in this direction.