Scene Information Research Articles

A Method for the Automatic Extraction of Support Devices in an Overhead Catenary System Based on MLS Point Clouds

A mobile laser scanning (MLS) system can acquire railway scene information quickly and provide a data foundation for regular railway inspections. The location of the catenary support device in an electrified railway system has a direct impact on the regular operation of the power supply system. However, multi-type support device data accounts for a tiny proportion of the whole railway scene, resulting in its poor characteristic expression in the scene. Therefore, using traditional point cloud filtering or point cloud segmentation methods alone makes it difficult to achieve an effective segmentation and extraction of the support device. As a result, this paper proposes an automatic extraction algorithm for complex railway support devices based on MLS point clouds. First, the algorithm obtains hierarchies of the pillar point clouds and the support device point clouds in the railway scene through high stratification and then realizes the noise that was point-cloud-filtered in the scene. Then, the center point of the pillar device is retrieved from the pillar corridor by a neighborhood search, and then the locating and initial extracting of the support device are realized based on the relatively stable spatial topological relationship between the pillar and the support device. Finally, a post-processing optimization method integrating the pillar filter and the voxelized projection filter is designed to achieve the accurate and efficient extraction of the support device based on the feature differences between the support device and other devices in the initial extraction results. Furthermore, in the experimental part, we evaluate the treatment effect of the algorithm in six types of support devices, three types of support device distribution scenes, and two types of railway units. The experimental results show that the average extraction IoU of the multi-type support device, support device distribution scenes, and railway unit were 97.20%, 94.29%, and 96.11%, respectively. In general, the proposed algorithm can achieve the accurate and efficient extraction of various support devices in different scenes, and the influence of the algorithm parameters on the extraction accuracy and efficiency is elaborated in the discussion section.

Visual object detection biases escape trajectories following acoustic startle in larval zebrafish

In this study, we investigated whether the larval zebrafish is sensitive to the presence of obstacles in its environment. Zebrafish execute fast escape swims when in danger of predation. We posited that collisions with solid objects during escape would be maladaptive to the fish, and therefore, the direction of escape swims should be informed by the locations of barriers. To test this idea, we developed a closed-loop imaging rig outfitted with barriers of various qualities. We show that when larval zebrafish escape in response to a non-directional vibrational stimulus, they use visual scene information to avoid collisions with obstacles. Our study demonstrates that barrier avoidance rate corresponds to the absolute distance of obstacles, as distant barriers outside of collision range elicit less bias than nearby collidable barriers that occupy the same amount of visual field. The computation of barrier avoidance is covert: the fact that fish will avoid barriers during escape cannot be predicted by its routine swimming behavior in the barrier arena. Finally, two-photon laser ablation experiments suggest that excitatory bias is provided to the Mauthner cell ipsilateral to approached barriers, either via direct excitation or a multi-step modulation process. We ultimately propose that zebrafish detect collidable objects via an integrative visual computation that is more complex than retinal occupancy alone, laying a groundwork for understanding how cognitive physical models observed in humans are implemented in an archetypal vertebrate brain. VIDEO ABSTRACT.

On the Co-Selection of Vision Transformer Features and Images for Very High-Resolution Image Scene Classification

Recent developments in remote sensing technology have allowed us to observe the Earth with very high-resolution (VHR) images. VHR imagery scene classification is a challenging problem in the field of remote sensing. Vision transformer (ViT) models have achieved breakthrough results in image recognition tasks. However, transformer–encoder layers encode different levels of features, where the latest layer represents semantic information, in contrast to the earliest layers, which contain more detailed data but ignore the semantic information of an image scene. In this paper, a new deep framework is proposed for VHR scene understanding by exploring the strengths of ViT features in a simple and effective way. First, pre-trained ViT models are used to extract informative features from the original VHR image scene, where the transformer–encoder layers are used to generate the feature descriptors of the input images. Second, we merged the obtained features as one signal data set. Third, some extracted ViT features do not describe well the image scenes, such as agriculture, meadows, and beaches, which could negatively affect the performance of the classification model. To deal with this challenge, we propose a new algorithm for feature- and image selection. Indeed, this gives us the possibility of eliminating the less important features and images, as well as those that are abnormal; based on the similarity of preserving the whole data set, we selected the most informative features and important images by dropping the irrelevant images that degraded the classification accuracy. The proposed method was tested on three VHR benchmarks. The experimental results demonstrate that the proposed method outperforms other state-of-the-art methods.

An Object Detection and Localization Method Based on Improved YOLOv5 for the Teleoperated Robot

In the traditional teleoperation system, the operator locates the object using the real-time scene information sent back from the robot terminal; however, the localization accuracy is poor and the execution efficiency is low. To address the issues, we propose an object detection and localization method for the teleoperated robot. First, we improved the classic YOLOv5 network model to produce superior object detection performance and named the improved model YOLOv5_Tel. On the basis of the classic YOLOv5 network model, the feature pyramid network was changed to a bidirectional feature pyramid network (BiFPN) network module to achieve the weighted feature fusion mechanism. The coordinate attention (CA) module was added to make the model pay more attention to the features of interest. Furthermore, we pruned the model from the depth and width to make it more lightweight and changed the bounding box regression loss function GIOU to SIOU to speed up model convergence. Then, the YOLOv5_Tel model and ZED2 depth camera were used to achieve object localization based on the binocular stereo vision ranging principle. Finally, we established an object detection platform for the teleoperated robot and created a small dataset to validate the proposed method. The experiment shows that compared with the classic YOLOv5 series network model, the YOLOv5_Tel is higher in accuracy, lighter in weight, and faster in detection speed. The mean average precision (mAP) value of the YOLOv5_Tel increased by 0.8%, 0.9%, and 1.0%, respectively. The model size decreased by 11.1%, 70.0%, and 86.4%, respectively. The inference time decreased by 9.1%, 42.9%, and 58.3%, respectively. The proposed object localization method has a high localization accuracy with an average relative error of only 1.12%.

Mapping Irregular Local Climate Zones from Sentinel-2 Images Using Deep Learning with Sequential Virtual Scenes

Recently, the local climate zone (LCZ) system has been presented to establish the connection between urban landscape and local thermal environment. However, LCZ entities are very difficult to be identified by pixel-based classifiers or object-oriented image analysis, as they are often a complicated combination of multiple ground objects (e.g., buildings, roads, grassland, etc.). Scene classifiers, especially deep learning methods can exploit the structure or contextual information of image scenes and then improve the performance of LCZ classification. However, the square and uniform-sized image patches often bring about extra challenges, as they cannot exactly match LCZ entities of diverse sizes and shapes in most cases. In this study, a sequential virtual scene method is presented to identify LCZ entities of diverse shapes and sizes, which consists of a small “core patch” for scanning diverse entities and sequential virtual scenes for providing abundant context. Specifically, the Bidirectional Long Short-Term Memory (Bi-LSTM) were used to learn the spatial relationship among virtual scenes, respectively. Importantly, a “self-attention” mechanism is designed to weigh the contribution of every virtual scene for alleviating the influences of mixed patches, according to the similarity between its hidden state and the final hidden state. Experiments prove SVS achieves better accuracies than random forest and ResNet and has the outstanding capacity of identifying irregular LCZ entities. It is a promising way to carry out LCZ mapping in cities of different types due to its flexibility and adaptability.

ST‐SIGMA: Spatio‐temporal semantics and interaction graph aggregation for multi‐agent perception and trajectory forecasting

Scene perception and trajectory forecasting are two fundamental challenges that are crucial to a safe and reliable autonomous driving (AD) system. However, most proposed methods aim at addressing one of the two challenges mentioned above with a single model. To tackle this dilemma, this paper proposes spatio-temporal semantics and interaction graph aggregation for multi-agent perception and trajectory forecasting (ST-SIGMA), an efficient end-to-end method to jointly and accurately perceive the AD environment and forecast the trajectories of the surrounding traffic agents within a unified framework. ST-SIGMA adopts a trident encoder–decoder architecture to learn scene semantics and agent interaction information on bird’s-eye view (BEV) maps simultaneously. Specifically, an iterative aggregation network is first employed as the scene semantic encoder (SSE) to learn diverse scene information. To preserve dynamic interactions of traffic agents, ST-SIGMA further exploits a spatio-temporal graph network as the graph interaction encoder. Meanwhile, a simple yet efficient feature fusion method to fuse semantic and interaction features into a unified feature space as the input to a novel hierarchical aggregation decoder for downstream prediction tasks is designed. Extensive experiments on the nuScenes data set have demonstrated that the proposed ST-SIGMA achieves significant improvements compared to the state-of-the-art (SOTA) methods in terms of scene perception and trajectory forecasting, respectively. Therefore, the proposed approach outperforms SOTA in terms of model generalisation and robustness and is therefore more feasible for deployment in real-world AD scenarios.

A multimodal deep neural network for prediction of the driver’s focus of attention based on anthropomorphic attention mechanism and prior knowledge

The prediction of the driver’s focus of attention (DFoA) is becoming essential research for the driver distraction detection and intelligent vehicle. Therefore, this work makes an attempt to predict DFoA. However, traffic driving environment is a complex and dynamic changing scene. The existing methods lack full utilization of driving scene information and ignore the importance of different objects or regions of the driving scene. To alleviate this, we propose a multimodal deep neural network based on anthropomorphic attention mechanism and prior knowledge (MDNN-AAM-PK). Specifically, a more comprehensive information of driving scene (RGB images, semantic images, optical flow images and depth images of successive frames) is as the input of MDNN-AAM-PK. An anthropomorphic attention mechanism is developed to calculate the importance of each pixel in the driving scene. A graph attention network is adopted to learn semantic context features. The convolutional long short-term memory network (ConvLSTM) is used to achieve the transition of fused features in successive frames. Furthermore, a training method based on prior knowledge is designed to improve the efficiency of training and the performance of DFoA prediction. These experiments, including experimental comparison with the state-of-the-art methods, the ablation study of the proposed method, the evaluation on different datasets and the visual assessment experiment in vehicle simulation platform, show that the proposed method can accurately predict DFoA and is better than the state-of-the-art methods.

Learning Moiré Pattern Elimination in Both Frequency and Spatial Domains for Image Demoiréing.

Recently, with the rapid development of mobile sensing technology, capturing scene information by mobile sensing devices in the form of images or videos has become a prevalent recording method. However, the moiré pattern phenomenon may occur when the scene contains digital screens or regular strips, which greatly degrade the visual performance and image quality. In this paper, considering the complexity and diversity of moiré patterns, we propose a novel end-to-end image demoiré method, which can learn moiré pattern elimination in both the frequency and spatial domains. To be specific, in the frequency domain, considering the signal energy of moiré pattern is widely distributed in the frequency, we introduce a wavelet transform to decompose the multi-scale image features, which can help the model identify the moiré features more precisely to suppress them effectively. On the other hand, we also design a spatial domain demoiré block (SDDB). The SDDB module can extract moiré features from the mixed features, then subtract them to obtain clean image features. The combination of the frequency domain and the spatial domain enhances the model's ability in terms of moiré feature recognition and elimination. Finally, extensive experiments demonstrate the superior performance of our proposed method to other state-of-the-art methods. The Grad-CAM results in our ablation study fully indicate the effectiveness of the two proposed blocks in our method.

Identification and Classification of Urban PLES Spatial Functions Based on Multisource Data and Machine Learning

Production space, living space, and ecological space (PLES) increasingly restrict and influence each other, and the urban PLES conflict significantly affects the sustainable development of a city. This study extracts multi-dimensional features from high-resolution remote sensing images, building vectors, points of interest (POI), and nighttime lighting data, and applies them to urban PLES feature recognition, dividing Ningbo into an agricultural production space, industrial and commercial production space, public living space, resident living space and ecological space. The specific research work was as follows: first, a convolutional neural network (CNN) was used to extract high-rise scene information from high-resolution remote sensing images; at the same time, through the geostatistical method, the building vector features, POI features, and night light features were extracted to express the economic and social characteristics of a city. Then, we used the nearest neighbor algorithm, decision-making tree algorithm, and random forest algorithm to train individual and combined features. Finally, random forest, which had the best training effect, was selected as the classifier in the fusion stage; as a result, the prediction accuracy rate reached 90.79%. The experimental results showed that the recognition model, based on multisource data and machine learning, had a good classification effect. Finally, we analyzed the current situation of the spatial distribution of PLES in Ningbo.

Transversal functional connectivity and scene-specific processing in the human entorhinal-hippocampal circuitry.

Scene and object information reach the entorhinal-hippocampal circuitry in partly segregated cortical processing streams. Converging evidence suggests that such information-specific streams organize the cortical - entorhinal interaction and the circuitry's inner communication along the transversal axis of hippocampal subiculum and CA1. Here, we leveraged ultra-high field functional imaging and advance Maass et al., 2015 who report two functional routes segregating the entorhinal cortex (EC) and the subiculum. We identify entorhinal subregions based on preferential functional connectivity with perirhinal Area 35 and 36, parahippocampal and retrosplenial cortical sources (referred to as ECArea35-based, ECArea36-based, ECPHC-based, ECRSC-based, respectively). Our data show specific scene processing in the functionally connected ECPHC-based and distal subiculum. Another route, that functionally connects the ECArea35-based and a newly identified ECRSC-based with the subiculum/CA1 border, however, shows no selectivity between object and scene conditions. Our results are consistent with transversal information-specific pathways in the human entorhinal-hippocampal circuitry, with anatomically organized convergence of cortical processing streams and a unique route for scene information. Our study thus further characterizes the functional organization of this circuitry and its information-specific role in memory function.

LFRSNet: A robust light field semantic segmentation network combining contextual and geometric features

Light field (LF) semantic segmentation is a newly arisen technology and is widely used in many smart city applications such as remote sensing, virtual reality and 3D photogrammetry. Compared with RGB images, LF images contain multi-layer contextual information and rich geometric information of real-world scenes, which are challenging to be fully exploited because of the complex and highly inter-twined structure of LF. In this paper, LF Contextual Feature (LFCF) and LF Geometric Feature (LFGF) are proposed respectively for occluded area perception and segmentation edge refinement. With exploitation of all the views in LF, LFCF provides glimpse of some occluded areas from other angular positions besides the superficial color information of the target view. The multi-layer information of the occluded area enhances the classification of partly occluded objects. Whereas LFGF is extracted from Ray Epipolar-Plane Images (RayEPIs) in eight directions for geometric information embedding. The solid geometric information refines object edges, especially for occlusion boundaries with similar colors. At last, Light Field Robust Segmentation Network (LFRSNet) is designed to integrate LFCF and LFGF. Multi-layer contextual information and geometric information are effectively incorporated through LFRSNet, which brings significant improvement for segmentation of the occluded objects and the object edges. Experimental results on both realworld and synthetic datasets proves the state-of-the-art performance of our method. Compared with other methods, LFRSNet produces more accurate segmentation under occlusion, especially in the edge regions.

Two-Stream Based Multi-Stage Hybrid Decoder for Self-Supervised Multi-Frame Monocular Depth

Self-supervised depth estimation has attracted a lot of attention recently due to its low cost. Despite using the self-supervision from image sequences, the current single-image based methods only infer depth from the scene information ignoring the matching information which is also important. Nevertheless, the matching information is not always reliable, especially in the texture-less and occlusion regions. Thus it would be attractive to combine the strength of single-image scene information and multi-frame matching information. In this letter, we propose a two-stream based multi-stage hybrid decoder to effectively accomplish the integration procedure. The hybrid decoder consists of two pathways for these two kinds of information respectively, and interactively fuses them. Specifically, a cost volume is built based on the scene prior to represent the matching information, and feeds back to the single-image pathway to complete the integration. To further facilitate the interactive integration, a multi-stage fusion strategy is embedded seamlessly into the hybrid decoder, resulting in more accurate depth results. Our approach outperforms the existing self-supervised methods on the KITTI and Cityscapes datasets.

Underwater multi-frame target images mosaic method based on adaptive image enhancement

The severe attenuation and scattering of light in the water reduces the effective field of view of the underwater camera, making the scene information contained in a single image limited, and it is difficult to meet the application requirements of the large-scale underwater scenes. To solve this problem, an underwater multi-frame target images mosaic method based on adaptive image enhancement is proposed in this paper. Firstly, the image blur prior is used to achieve the adaptive enhancement of underwater images to suppress the blur and colour distortion of underwater images. Then, the feature point matching method based on the improved SURF realizes the extraction and matching of the feature points of underwater multi-frame target images. Finally, combining with the fusion strategy of gradual in and out, the seamless splicing and fusion of underwater multi-frame target images is realized. The pool and shallow sea tests were carried out respectively, and the results show that the method proposed in this paper increases the number of effective feature point matching and improves the splicing effect.

The human posterior cingulate, retrosplenial, and medial parietal cortex effective connectome, and implications for memory and navigation.

The human posterior cingulate, retrosplenial, and medial parietal cortex are involved in memory and navigation. The functional anatomy underlying these cognitive functions was investigated by measuring the effective connectivity of these Posterior Cingulate Division (PCD) regions in the Human Connectome Project-MMP1 atlas in 171 HCP participants, and complemented with functional connectivity and diffusion tractography. First, the postero-ventral parts of the PCD (31pd, 31pv, 7m, d23ab, and v23ab) have effective connectivity with the temporal pole, inferior temporal visual cortex, cortex in the superior temporal sulcus implicated in auditory and semantic processing, with the reward-related vmPFC and pregenual anterior cingulate cortex, with the inferior parietal cortex, and with the hippocampal system. This connectivity implicates it in hippocampal episodic memory, providing routes for "what," reward and semantic schema-related information to access the hippocampus. Second, the antero-dorsal parts of the PCD (especially 31a and 23d, PCV, and also RSC) have connectivity with early visual cortical areas including those that represent spatial scenes, with the superior parietal cortex, with the pregenual anterior cingulate cortex, and with the hippocampal system. This connectivity implicates it in the "where" component for hippocampal episodic memory and for spatial navigation. The dorsal-transitional-visual (DVT) and ProStriate regions where the retrosplenial scene area is located have connectivity from early visual cortical areas to the parahippocampal scene area, providing a ventromedial route for spatial scene information to reach the hippocampus. These connectivities provide important routes for "what," reward, and "where" scene-related information for human hippocampal episodic memory and navigation. The midcingulate cortex provides a route from the anterior dorsal parts of the PCD and the supracallosal part of the anterior cingulate cortex to premotor regions.

Underwater Holothurian Target-Detection Algorithm Based on Improved CenterNet and Scene Feature Fusion.

Aiming at the common problems, such as noise pollution, low contrast, and color distortion in underwater images, and the characteristics of holothurian recognition, such as morphological ambiguity, high similarity with the background, and coexistence of special ecological scenes, this paper proposes an underwater holothurian target-detection algorithm (FA-CenterNet), based on improved CenterNet and scene feature fusion. First, to reduce the model’s occupancy of embedded device resources, we use EfficientNet-B3 as the backbone network to reduce the model’s Params and FLOPs. At the same time, EfficientNet-B3 increases the depth and width of the model, which improves the accuracy of the model. Then, we design an effective FPT (feature pyramid transformer) combination module to fully focus and mine the information on holothurian ecological scenarios of different scales and spaces (e.g., holothurian spines, reefs, and waterweeds are often present in the same scenario as holothurians). The co-existing scene information can be used as auxiliary features to detect holothurians, which can improve the detection ability of fuzzy and small-sized holothurians. Finally, we add the AFF module to realize the deep fusion of the shallow-detail and high-level semantic features of holothurians. The results show that the method presented in this paper yields better results on the 2020 CURPC underwater target-detection image dataset with an AP50 of 83.43%, Params of 15.90 M, and FLOPs of 25.12 G compared to other methods. In the underwater holothurian-detection task, this method improves the accuracy of detecting holothurians with fuzzy features, a small size, and dense scene. It also achieves a good balance between detection accuracy, Params, and FLOPs, and is suitable for underwater holothurian detection in most situations.

Spatiotemporal Feature Enhancement Aids the Driving Intention Inference of Intelligent Vehicles.

In order that fully self-driving vehicles can be realized, it is believed that systems where the driver shares control and authority with the intelligent vehicle offer the most effective solution. An understanding of driving intention is the key to building a collaborative autonomous driving system. In this study, the proposed method incorporates the spatiotemporal features of driver behavior and forward-facing traffic scenes through a feature extraction module; the joint representation was input into an inference module for obtaining driver intentions. The feature extraction module was a two-stream structure that was designed based on a deep three-dimensional convolutional neural network. To accommodate the differences in video data inside and outside the cab, the two-stream network consists of a slow pathway that processes the driver behavior data with low frame rates, along with a fast pathway that processes traffic scene data with high frame rates. Then, a gated recurrent unit, based on a recurrent neural network, and a fully connected layer constitute an intent inference module to estimate the driver’s lane-change and turning intentions. A public dataset, Brain4Cars, was used to validate the proposed method. The results showed that compared with modeling using the data related to driver behaviors, the ability of intention inference is significantly improved after integrating traffic scene information. The overall accuracy of the intention inference of five intents was 84.92% at a time of 1 s prior to the maneuver, indicating that making full use of traffic scene information was an effective way to improve inference performance.

Skeleton-based deep pose feature learning for action quality assessment on figure skating videos

Most of the existing Action Quality Assessment (AQA) methods for scoring sports videos have deeply researched how to evaluate the single action or several sequential-defined actions that performed in short-term sport videos, such as diving, vault, etc. They attempted to extract features directly from RGB videos through 3D ConvNets, which makes the features mixed with ambiguous scene information. To investigate the effectiveness of deep pose feature learning on automatically evaluating the complicated activities in long-duration sports videos, such as figure skating and artistic gymnastic, we propose a skeleton-based deep pose feature learning method to address this problem. For pose feature extraction, a spatial–temporal pose extraction module (STPE) is built to capture the subtle changes of human body movements and obtain the detail representations for skeletal data in space and time dimensions. For temporal information representation, an inter-action temporal relation extraction module (ATRE) is implemented by recurrent neural network to model the dynamic temporal structure of skeletal subsequences. We evaluate the proposed method on figure skating activity of MIT-skate and FIS-V datasets. The experimental results show that the proposed method is more effective than RGB video-based deep feature learning methods, including SENet and C3D. Significant performance progress has been achieved for the Spearman Rank Correlation (SRC) on MIT-Skate dataset. On FIS-V dataset, for the Total Element Score (TES) and the Program Component Score (PCS), better SRC and MSE have been achieved between the predicted scores against the judge’s ones when compared with SENet and C3D feature methods.

Absolute Camera Pose Regression Using an RGB-D Dual-Stream Network and Handcrafted Base Poses.

Absolute pose regression (APR) for camera localization is a single-shot approach that encodes the information of a 3D scene in an end-to-end neural network. The camera pose result of APR methods can be observed as the linear combination of the base poses. Previous APR methods’ base poses are learned from training data. However, the training data can limit the performance of the methods, which cannot be generalized to cover the entire scene. To solve this issue, we use handcrafted base poses instead of learning-based base poses, which prevents overfitting the camera poses of the training data. Moreover, we use a dual-stream network architecture to process color and depth images separately to get more accurate localization. On the 7 Scenes dataset, the proposed method is among the best in median rotation error, and in median translation error, it outperforms previous APR methods. On a more difficult dataset—Oxford RobotCar dataset, the proposed method achieves notable improvements in median translation and rotation errors compared to the state-of-the-art APR methods.

3D semantic segmentation based on spatial-aware convolution and shape completion for augmented reality applications

3D semantic segmentation of indoor scenes is a popular research topic in the field of computer vision. For many applications, it is very important to know exactly what category each point in the scene belongs to. Benefiting from the development of deep learning, many neural networks based on voxels and points have been proposed to solve these segmentation problems. However, most of them do not fully consider the information of the spatial structure. Current voxel-based sparse convolutional neural networks can effectively extract 3D features in space. However, they assume that the feature in the empty space is zero, causing a loss of information in the spatial structure. In this paper, we propose a system that uses scene point clouds with color information to semantically segment an entire indoor scene. Based on the sparsity of spatial data, we design a novel spatial-aware sparse convolution operation. We encode the spatial information of the object’s existence as an additional feature and use the self-attention mechanism to effectively aggregate features. In addition, we introduce a completion network to refine the results from the segmentation network, so that each object in the scene is fitted into a more reasonable and complete shape. Through the above two methods, we build an accurate scene semantic segmentation network to obtain the semantic information of the entire scene. In the experimental part, we use two public datasets to perform quantitative and qualitative analysis. We compare our results with other state-of-the-art methods to prove the superiority of our method. Our models are also examined under different configurations to assure the effectiveness of the proposed method. Finally, the semantic segmentation model was integrated into a real-world application to demonstrate its usefulness. We expect that the proposed 3D scene semantic segmentation system can provide accurate and fast results for practical applications.

Saliency and Granularity: Discovering Temporal Coherence for Video-Based Person Re-Identification

Video-based person re-identification (ReID) matches the same people across the video sequences with rich spatial and temporal information in complex scenes. It is highly challenging to capture discriminative information when occlusions and pose variations exist between frames. A key solution to this problem rests on extracting the temporal invariant features of video sequences. In this paper, we propose a novel method for discovering temporal coherence by designing a region-level saliency and granularity mining network (SGMN). Firstly, to address the varying noisy frame problem, we design a temporal spatial-relation module (TSRM) to locate frame-level salient regions, adaptively modeling the temporal relations on spatial dimension through a probe-buffer mechanism. It avoids the information redundancy between frames and captures the informative cues of each frame. Secondly, a temporal channel-relation module (TCRM) is proposed to further mine the small granularity information of each frame, which is complementary to TSRM by concentrating on discriminative small-scale regions. TCRM exploits a one-and-rest difference relation on channel dimension to enhance the granularity features, leading to stronger robustness against misalignments. Finally, we evaluate our SGMN with four representative video-based datasets, including iLIDS-VID, MARS, DukeMTMC-VideoReID, and LS-VID, and the results indicate the effectiveness of the proposed method.