Knowledge distillation (KD) transfers the ``dark knowledge'' from a complex teacher model to a compact student model. However, heterogeneous architecture distillation, such as Vision Transformer (ViT) to ResNet18, faces challenges due to differences in spatial feature representations. Traditional KD methods are mostly designed for homogeneous architectures and hence struggle to effectively address the disparity. Although heterogeneous KD approaches have been developed recently to solve these issues, they often incur high computational costs and complex designs, or overly rely on logit alignment, which limits their ability to leverage the complementary features. To overcome these limitations, we propose Heterogeneous Complementary Distillation (HCD), a simple yet effective framework that integrates complementary teacher and student features to align representations in shared logits. These logits are decomposed and constrained to facilitate diverse knowledge transfer to the student. Specifically, HCD processes the student’s intermediate features through convolutional projector and adaptive pooling, concatenates them with teacher's feature from the penultimate layer and then maps them via the Complementary Feature Mapper (CFM) module, comprising fully connected layer, to produce shared logits. We further introduce Sub-logit Decoupled Distillation (SDD) that partitions the shared logits into n sub-logits, which are fused with teacher's logits to rectify classification. To ensure sub-logit diversity and reduce redundant knowledge transfer, we propose an Orthogonality Loss (OL). By preserving student-specific strengths and leveraging teacher knowledge, HCD enhances robustness and generalization in students. Extensive experiments on the CIFAR-100, fine-grained (e.g., CUB200, Aircraft) and ImageNet-1K datasets demonstrate that HCD outperforms state-of-the-art KD methods, establishing it as an effective solution for heterogeneous KD.

Sharding, as a Layer-1 scaling technique, is widely recognized as a promising solution to address the scalability limitations faced by blockchains. However, distributing accounts across shards generates numerous cross-shard transactions (CSTs) and inter-shard workload imbalances, potentially compromising scalability. Existing state-of-the-art methods typically optimize transaction distribution for subsequent epochs by periodically reallocating accounts using graph partitioning or community detection to balance loads and minimize CSTs. Nevertheless, these methods often involve computationally expensive account migration and overlook real-world transaction skewness, exacerbating workload imbalances. To address the above problems, we propose a two-phase account group migration service with dynamic load awareness in this paper. This service can monitor and analyze system state to determine the necessity of account migration. Once triggered, it employs a two-phase algorithm that combines account grouping with global group-level re-partitioning to balance workloads and minimize CSTs. Additionally, group-level migration further helps reduce computational overhead. We evaluate the designed service by replaying large-scale real Ethereum transactions. Experimental results demonstrate that, compared with the baselines, our method can not only improves system throughput and reduces transaction confirmation latency, but also enhances overall scalability.

Knowledge distillation transfers "dark knowledge" from a large teacher model to a smaller student model, yielding a highly efficient network. To improve network's generalization ability, existing works use a larger temperature coefficient for knowledge distillation. Nevertheless, these methods may lower the target category's confidence and lead to ambiguous recognition of similar samples. To mitigate this issue, some studies introduce intra-batch distillation to reduce prediction discrepancy. However, these methods overlook the inconsistency between background information and the target category, which may increase prediction bias due to noise disturbance. Additionally, label imbalance from random sampling and batch size can undermine network generalization reliability. To tackle these challenges, we propose a simple yet effective Intra-class Knowledge Distillation (IKD) method that facilitates knowledge sharing within the same class to ensure consistent predictions. First, we initialize the matrix and the vector to store logits and class counts provided by the teacher, respectively. Then, in the first epoch, we calculate the sum of logits and sample counts per class and perform KD to prevent knowledge omission. Finally, in subsequent training, we update the matrix to obtain the average logits and compute the KL divergence between the student's output and the updated matrix according to the label index. This process ensures intra-class consistency and improves the student's performance. Furthermore, this method theoretically reduces prediction bias by ensuring intra-class consistency. Extensive experiments on the CIFAR-100, ImageNet-1K, and Tiny-ImageNet datasets validate the superiority of IKD.

State-of-the-art logit distillation methods exhibit versatility, simplicity, and efficiency. Despite the advances, existing studies have yet to delve thoroughly into fine-grained relationships within logit knowledge. In this paper, we propose Local Dense Relational Logit Distillation (LDRLD), a novel method that captures inter-class relationships through recursively decoupling and recombining logit information, thereby providing more detailed and clearer insights for student learning. To further optimize the performance, we introduce an Adaptive Decay Weight (ADW) strategy, which can dynamically adjust the weights for critical category pairs using Inverse Rank Weighting (IRW) and Exponential Rank Decay (ERD). Specifically, IRW assigns weights inversely proportional to the rank differences between pairs, while ERD adaptively controls weight decay based on total ranking scores of category pairs. Furthermore, after the recursive decoupling, we distill the remaining non-target knowledge to ensure knowledge completeness and enhance performance. Ultimately, our method improves the student's performance by transferring fine-grained knowledge and emphasizing the most critical relationships. Extensive experiments on datasets such as CIFAR-100, ImageNet-1K, and Tiny-ImageNet demonstrate that our method compares favorably with state-of-the-art logit-based distillation approaches. The code will be made publicly available.

Since rain shows a variety of shapes and directions, learning the degradation representation is extremely challenging for single image deraining. Existing methods mainly propose to designing complicated modules to implicitly learn latent degradation representation from rainy images. However, it is hard to decouple the content-independent degradation representation due to the lack of explicit constraint, resulting in over- or under-enhancement problems. To tackle this issue, we propose a novel Latent Degradation Representation Constraint Network (LDRCNet) that consists of the Direction-Aware Encoder (DAEncoder), Deraining Network, and Multi-Scale Interaction Block (MSIBlock). Specifically, the DAEncoder is proposed to extract latent degradation representation adaptively by first using the deformable convolutions to exploit the direction property of rain streaks. Next, a constraint loss is introduced to explicitly constraint the degradation representation learning during training. Last, we propose an MSIBlock to fuse with the learned degradation representation and decoder features of the deraining network for adaptive information interaction to remove various complicated rainy patterns and reconstruct image details. Experimental results on five synthetic and four real datasets demonstrate that our method achieves state-of-the-art performance. The source code is available at https://github.com/Madeline-hyh/LDRCNet.

Transformers have recently emerged as a significant force in the field of image deraining. Existing image deraining methods utilize extensive research on self-attention. Though showcasing impressive results, they tend to neglect critical frequency information, as self-attention is generally less adept at capturing high-frequency details. To overcome this shortcoming, we have developed an innovative Dual-Path Coupled Deraining Network (DPCNet) that integrates information from both spatial and frequency domains through Spatial Feature Extraction Block (SFEBlock) and Frequency Feature Extraction Block (FFEBlock). We have further introduced an effective Adaptive Fusion Module (AFM) for the dual-path feature aggregation. Extensive experiments on six public deraining benchmarks and downstream vision tasks have demonstrated that our proposed method not only outperforms the existing state-of-the-art deraining method but also achieves visually pleasuring results with excellent robustness on downstream vision tasks. The source code is available at https://github.com/Madeline-hyh/DPCNet.

High-performance real-time 3D object detection is crucial in autonomous driving perception systems. Voxel-or point-based 3D object detectors are highly accurate but inefficient and difficult to deploy, while other methods use 2D projection views to improve efficiency, but information loss usually degrades performance. To balance effectiveness and efficiency, we propose a scheme called RFDNet that uses range features to decorate points. Specifically, RFDNet adaptively aggregates point features projected to independent grids and nearby regions via Dilated Grid Feature Encoding (DGFE) to generate a range view, which can handle occlusion and multi-frame inputs while the established geometric correlation between grid with surrounding space weakens the effects of scale distortion. We also propose a Soft Box Regression (SBR) strategy that supervises 3D box regression on a more extensive range than conventional methods to enhance model robustness. In addition, RFDNet benefits from our designed Semantic-assisted Ground-truth Sample (SA-GTS) data augmentation, which additionally considers collisions and spatial distributions of objects. Experiments on the nuScenes benchmark show that RFDNet outperforms all LiDAR-only non-ensemble 3D object detectors and runs at high speed of 20 FPS, achieving a better effectiveness-efficiency trade-off. Code is available at https://github.com/wy17646051/RFDNet.

Abstract Despite their great advancement, current pedestrian detection methods focus on single static images, which fail to employ richer information available from the video sequences. Compared with still images, videos can offer temporal information of objects in the time dimension, thus providing the potential to obtain more robust detection performance. Here, a novel pedestrian detection method based on visible part detection and temporal calibration is proposed. Specifically, a part‐aware module to predict the visible body part of each pedestrian instance, which enables us to obtain precise motion information of partially occluded pedestrians in a video sequence, is first developed. Then, the temporal coherence for each pedestrian instance based on the predicted motion information is constructed. After that, an adaptive temporal calibration method is introduced to effectively calibrate the final detection result. This method on two video pedestrian detection benchmarks, that is, Caltech‐New and MOT17Det, is evaluated. Experimental results show that this method performs favourably against existing pedestrian detection approaches.

Automatic coronary artery segmentation is of great value in diagnosing coronary disease. In this paper, we propose an automatic coronary artery segmentation method for coronary computerized tomography angiography (CCTA) images based on a deep convolutional neural network. The proposed method consists of three steps. First, to improve the efficiency and effectiveness of the segmentation, a 2D DenseNet classification network is utilized to screen out the non-coronary-artery slices. Second, we propose a coronary artery segmentation network based on the 3D-UNet, which is capable of extracting, fusing and rectifying features efficiently for accurate coronary artery segmentation. Specifically, in the encoding process of the 3D-UNet network, we adapt the dense block into the 3D-UNet so that it can extract rich and representative features for coronary artery segmentation; In the decoding process, 3D residual blocks with feature rectification capability are applied to improve the segmentation quality further. Third, we introduce a Gaussian weighting method to obtain the final segmentation results. This operation can highlight the more reliable segmentation results at the center of the 3D data blocks while weakening the less reliable segmentations at the block boundary when merging the segmentation results of spatially overlapping data blocks. Experiments demonstrate that our proposed method achieves a Dice Similarity Coefficient (DSC) value of 0.826 on a CCTA dataset constructed by us. The code of the proposed method is available at https://github.com/alongsong/3D_CAS.

Pedestrian detection is a crucial task in intelligent transportation systems, which can be applied in autonomous vehicles and traffic scene video surveillance systems. The past few years have witnessed much progress on the research of pedestrian detection methods, especially through the successful use of the deep learning based techniques. However, occlusion and large scale variation remain the challenging issues for pedestrian detection. In this work, we propose a Part-Aware Multi-Scale Fully Convolutional Network (PAMS-FCN) to tackle these difficulties. Specifically, we present a part-aware Region-of-Interest (RoI) pooling module to mine body parts with different responses, and select the part with the strongest response via voting. As such, a partially visible pedestrian instance can receive a high detection confidence score, making it less likely to become a missing detection. This module operates in parallel with an instance RoI pooling module to combine local parts and global context information. To handle vast scale variation, we construct a fully convolutional network in which multi-scale feature maps are generated efficiently, and small-scale and large-scale pedestrians are detected separately. By integrating these structures, the proposed detector achieves the state-of-the-art performance on the Caltech, KITTI, INRIA and ETH pedestrian detection datasets.

Analysis of the morphology and function of the right ventricle (RV) can be used for the prediction and diagnosis of cardiovascular disease. Accurate description of the structure and function of heart can be provided by analyzing cardiac magnetic resonance imaging (MRI) images. Noise interference and intensity inhomogeneity of MRI images can be addressed by using a local intensity clustering (LIC) model. However, the segmentation of the RV in MRI images still remains a challenge mainly due to its ill-defined borders. To address such a challenge, an algorithm for segmenting the RV based on a local motion intensity clustering (LMIC) model is proposed in this paper. The LMIC model combines the LIC model with the motion intensity information, due to cardiac motion and blood flow. The motion intensity is calculated by using the Lucas Kanade optical flow method and utilized in the LMIC model as an energy parameter. Because the motion intensity of the RV region is stronger than other areas, the RV can be accurately segmented by this approach. Experimental results demonstrate that the LMIC model is able to address the challenge of the ill-defined RV borders in cardiac MRI images and improved RV segmentation accuracy over existing methods.

We propose an online multiple object tracking algorithm that exploits optical flow and convolutional features to handle noisy detections as well as frequent occlusion. To achieve robust tracking, we develop a data association method that deals with tracking scenarios of increasing difficulty. For easy scenarios, we use motion affinity to associate detections with objects. For ambiguous situations, we propose to use an appearance model based on convolutional features and correlation filters to complement template matching methods. For difficult cases where objects are under heavy occlusion, we carry out occlusion analysis, which exploits the relationship between targets and occluders to predict potential object locations. To deal with noisy detections, false positives are detected and removed on both raw detection and tracklet levels, while missing and inaccurate detections are recovered or corrected via short-term tracking. Experimental results on two benchmark datasets demonstrate that the proposed online algorithm performs favorably against the state-of-the-art methods.

Despite significant progress in pedestrian detection has been made in recent years, detecting pedestrians in crowded scenes remains a challenging problem. In this paper, we propose to use visual contexts based on scale and occlusion cues from detections at proximity to better detect pedestrians for surveillance applications. Specifically, we first apply detectors based on full body and parts to generate initial detections. Scale prior at each image location is estimated using the cues provided by neighboring detections, and the confidence score of each detection is refined according to its consistency with the estimated scale prior. Local occlusion analysis is exploited in refining detection confidence scores which facilitates the final detection cluster based Non-Maximum Suppression. Experimental results on benchmark data sets show that the proposed algorithm performs favorably against the state-of-the-art methods.

Multiple-object tracking is an important task in automated video surveillance. In this paper, we present a multiple-human-tracking approach that takes the single-frame human detection results as input and associates them to form trajectories while improving the original detection results by making use of reliable temporal information in a closed-loop manner. It works by first forming tracklets, from which reliable temporal information is extracted, and then refining the detection responses inside the tracklets, which also improves the accuracy of tracklets' quantities. After this, local conservative tracklet association is performed and reliable temporal information is propagated across tracklets so that more detection responses can be refined. The global tracklet association is done last to resolve association ambiguities. Experimental results show that the proposed approach improves both the association and detection results. Comparison with several state-of-the-art approaches demonstrates the effectiveness of the proposed approach.

In this paper, the problem of human detection in crowded scenes is formulated as a maximum a posteriori problem, in which, given a set of candidates, predefined 3-D human shape models are matched with image evidence, provided by foreground extraction and probability of boundary, to estimate the human configuration. The optimal solution is obtained by decomposing the mutually related candidates into unoccluded and occluded ones in each iteration according to a graph description of the candidate relations and then only matching models for the unoccluded candidates. A candidate validation and rejection process based on minimum description length and local occlusion reasoning is carried out after each iteration of model matching. The advantage of the proposed optimization procedure is that its computational cost is much smaller than that of global optimization methods, while its performance is comparable to them. The proposed method achieves a detection rate of about 2% higher on a subset of images of the Caviar data set than the best result reported by previous works. We also demonstrate the performance of the proposed method using another challenging data set.

In this paper, we solve the problem of human detection in crowded scenes using a Bayesian 3D model based method. Human candidates are first nominated by a head detector and a foot detector, then optimization is performed to find the best configuration of the candidates and their corresponding shape models. The solution is obtained by decomposing the mutually related candidates into un-occluded ones and occluded ones in each iteration, and then performing model matching for the un-occluded candidates. To this end, in addition to some obvious clues, we also derive a graph that depicts the inter-object relation so that unreasonable decomposition is avoided. The merit of the proposed optimization procedure is that its computational cost is similar to the greedy optimization methods while its performance is comparable to the global optimization approaches. For model matching, it is performed by employing both prior knowledge and image likelihood, where the priors include the distribution of individual shape models and the restriction on the inter-object distance in real world, and image likelihood is provided by foreground extraction and the edge information. After the model matching, a validation and rejection strategy based on minimum description length is applied to confirm the candidates that have reliable matching results. The proposed method is tested on both the publicly available Caviar dataset and a challenging dataset constructed by ourselves. The experimental results demonstrate the effectiveness of our approach.

Tracking by association of low frame rate detection responses is not trivial, as motion is less continuous and hence ambiguous. The problem becomes more challenging when occlusion occurs. To solve this problem, we firstly propose a robust data association method that explicitly differentiates ambiguous tracklets that are likely to introduce incorrect linking from other tracklets, and deal with them effectively. Secondly, we solve the long-time occlusion problem by detecting inter-track relationship and performing track split and merge according to appearance similarity and occlusion order. Experiment on a challenging human surveillance dataset shows the effectiveness of the proposed method.

The extraction of moving objects from their background is a challenging task in visual surveillance. As a single threshold often fails to resolve ambiguities and correctly segment the object, in this paper, we propose a new method that uses three thresholds to accurately classify pixels as foreground or background. These thresholds are adaptively determined by considering the distributions of differences between the input and background images and are used to generate three boundary sets. These boundary sets are then merged to produce a final boundary set that represents the boundaries of the moving objects. The merging step proceeds by first identifying boundary segment pairs that are significantly inconsistent. Then, for each inconsistent boundary segment pair, its associated curvature, edge response, and shadow index are used as criteria to evaluate the probable location of the true boundary. The resulting boundary is finally refined by estimating the width of the halo-like boundary and referring to the foreground edge map. Experimental results show that the proposed method consistently performs well under different illumination conditions, including indoor, outdoor, moderate, sunny, rainy, and dim cases. By comparing with a ground truth in each case, both the classification error rate and the displacement error indicate an accurate detection, which show substantial improvement in comparison with other existing methods.

In this paper, the crowd counting and segmentation problem is formulated as a maximum a posterior problem, in which 3D human shape models are designed and matched with image evidence provided by foreground/background separation and probability of boundary. The solution is obtained by considering only the human candidates that are possible to be un-occluded in each iteration, and then applying on them a validation and rejection strategy based on minimum description length. The merit of the proposed optimization procedure is that its computational cost is much smaller than that of the global optimization methods while its performance is comparable to them. The approach is shown to be robust with respect to severe partial occlusions.

This paper compares the radial artery pulses of 105 young graduate students. The radial artery pulses after performing progressive ergometer for five minutes are different from those at rest. All the pulses become floating and fast. The contours of pulses have three kinds of variability. The incisures of 39 subjects become especially low; sometimes the incisures are lower than the onset of pulse waveform. The tidal waves and dicrotic waves of 32 subjects become higher. The pulses of 34 subjects become smooth. Their incisures and dicrotic waves become lower. These changes can instruct the exercise and training of the young students and athletes