Human activity recognition(HAR)for dense prediction is proven to be of good performance,but it relies on labeling every point in time series with the high cost.In addition,the performance of HAR model will show signif...Human activity recognition(HAR)for dense prediction is proven to be of good performance,but it relies on labeling every point in time series with the high cost.In addition,the performance of HAR model will show significant degradation when tested on the sensor data with different distribution from the training data,where the training data and the test data are usually collected from different sensor locations or sensor users.Therefore,the adaptive transfer learning framework for dense prediction of HAR is introduced to implement cross-domain transfer,where the proposed multi-level unsupervised domain adaptation(MLUDA)approach combines the global domain adaptation and the specific task adaptation to adapt the source and target domain in multiple levels.The multi-connected global domain adaptation architecture is proposed for the first time,which can adapt the output layer of the encoder and the decoder in dense prediction model.After this,the specific task adaptation is proposed to ensure alignment of each class centroid in source domain and target domain by introducing the cosine distance loss and the moving average method.Experiments on three public HAR datasets demonstrate that the proposed MLUDA improves the prediction accuracy of target data by 20%compared to the source domain pre-trained model and it is more effective than the other three deep transfer learning methods with an improvement of 10%to 18%in accuracy.展开更多
Bubble size distribution is the basic apparent performance and obvious characteristics in the air dense medium fluidized bed (ADMFB). The approaches of numerical simulation and experimental verification were combined ...Bubble size distribution is the basic apparent performance and obvious characteristics in the air dense medium fluidized bed (ADMFB). The approaches of numerical simulation and experimental verification were combined to conduct the further research on the bubble generation and movement behavior. The results show that ADMFB could display favorable expanded characteristics after steady fluidization. With different particle size distributions of magnetite powder as medium solids, we selected an appropriate prediction model for the mean bubble diameter in ADMFB. The comparison results indicate that the mean bubble diameters along the bed heights are 35 mm < D b < 66 mm and 40 mm < D b < 69 mm with the magnetite powder of 0.3 mm+0.15mm and 0.15mm+0.074mm, respectively. The prediction model provides good agreements with the experimental and simulation data. Based on the optimal operating gas velocity distribution, the mixture of magnetite powder and <1mm fine coal as medium solids were utilized to carry out the separation experiment on 6-50mm raw coal. The results show that an optimal separation density d P of 1.73g/cm 3 with a probable error E of 0.07g/cm 3 and a recovery efficiency of 99.97% is achieved, which indicates good separation performance by applying ADMFB.展开更多
Self-supervised learning aims to learn a universal feature representation without labels.To date,most existing self-supervised learning methods are designed and optimized for image classification.These pre-trained mod...Self-supervised learning aims to learn a universal feature representation without labels.To date,most existing self-supervised learning methods are designed and optimized for image classification.These pre-trained models can be sub-optimal for dense prediction tasks due to the discrepancy between image-level prediction and pixel-level prediction.To fill this gap,we aim to design an effective,dense self-supervised learning framework that directly works at the level of pixels(or local features)by taking into account the correspondence between local features.Specifically,we present dense contrastive learning(DenseCL),which implements self-supervised learning by optimizing a pairwise contrastive(dis)similarity loss at the pixel level between two views of input images.Compared to the supervised ImageNet pre-training and other self-supervised learning methods,our self-supervised DenseCL pretraining demonstrates consistently superior performance when transferring to downstream dense prediction tasks including object detection,semantic segmentation and instance segmentation.Specifically,our approach significantly outperforms the strong MoCo-v2 by 2.0%AP on PASCAL VOC object detection,1.1%AP on COCO object detection,0.9%AP on COCO instance segmentation,3.0%mIoU on PASCAL VOC semantic segmentation and 1.8%mIoU on Cityscapes semantic segmentation.The improvements are up to 3.5%AP and 8.8%mIoU over MoCo-v2,and 6.1%AP and 6.1%mIoU over supervised counterpart with frozen-backbone evaluation protocol.展开更多
Transformers have recently lead to encouraging progress in computer vision.In this work,we present new baselines by improving the original Pyramid Vision Transformer(PVT v1)by adding three designs:(i)a linear complexi...Transformers have recently lead to encouraging progress in computer vision.In this work,we present new baselines by improving the original Pyramid Vision Transformer(PVT v1)by adding three designs:(i)a linear complexity attention layer,(ii)an overlapping patch embedding,and(iii)a convolutional feed-forward network.With these modifications,PVT v2 reduces the computational complexity of PVT v1 to linearity and provides significant improvements on fundamental vision tasks such as classification,detection,and segmentation.In particular,PVT v2 achieves comparable or better performance than recent work such as the Swin transformer.We hope this work will facilitate state-ofthe-art transformer research in computer vision.Code is available at https://github.com/whai362/PVT.展开更多
基金supported by the State Major Science and Technology Special Projects (2014ZX03004002)Fab. X Artificial Intelligence Research Center,Beijing,P. R. C.
文摘Human activity recognition(HAR)for dense prediction is proven to be of good performance,but it relies on labeling every point in time series with the high cost.In addition,the performance of HAR model will show significant degradation when tested on the sensor data with different distribution from the training data,where the training data and the test data are usually collected from different sensor locations or sensor users.Therefore,the adaptive transfer learning framework for dense prediction of HAR is introduced to implement cross-domain transfer,where the proposed multi-level unsupervised domain adaptation(MLUDA)approach combines the global domain adaptation and the specific task adaptation to adapt the source and target domain in multiple levels.The multi-connected global domain adaptation architecture is proposed for the first time,which can adapt the output layer of the encoder and the decoder in dense prediction model.After this,the specific task adaptation is proposed to ensure alignment of each class centroid in source domain and target domain by introducing the cosine distance loss and the moving average method.Experiments on three public HAR datasets demonstrate that the proposed MLUDA improves the prediction accuracy of target data by 20%compared to the source domain pre-trained model and it is more effective than the other three deep transfer learning methods with an improvement of 10%to 18%in accuracy.
基金financially supported by the National Natural Science Foundation of China (Nos. 51221462, 51134022,51174203 and 51074156)the National Basic Research Program of China (No. 2012CB214904)China Postdoctoral Science Foundation (No. 2013M531430)
文摘Bubble size distribution is the basic apparent performance and obvious characteristics in the air dense medium fluidized bed (ADMFB). The approaches of numerical simulation and experimental verification were combined to conduct the further research on the bubble generation and movement behavior. The results show that ADMFB could display favorable expanded characteristics after steady fluidization. With different particle size distributions of magnetite powder as medium solids, we selected an appropriate prediction model for the mean bubble diameter in ADMFB. The comparison results indicate that the mean bubble diameters along the bed heights are 35 mm < D b < 66 mm and 40 mm < D b < 69 mm with the magnetite powder of 0.3 mm+0.15mm and 0.15mm+0.074mm, respectively. The prediction model provides good agreements with the experimental and simulation data. Based on the optimal operating gas velocity distribution, the mixture of magnetite powder and <1mm fine coal as medium solids were utilized to carry out the separation experiment on 6-50mm raw coal. The results show that an optimal separation density d P of 1.73g/cm 3 with a probable error E of 0.07g/cm 3 and a recovery efficiency of 99.97% is achieved, which indicates good separation performance by applying ADMFB.
文摘Self-supervised learning aims to learn a universal feature representation without labels.To date,most existing self-supervised learning methods are designed and optimized for image classification.These pre-trained models can be sub-optimal for dense prediction tasks due to the discrepancy between image-level prediction and pixel-level prediction.To fill this gap,we aim to design an effective,dense self-supervised learning framework that directly works at the level of pixels(or local features)by taking into account the correspondence between local features.Specifically,we present dense contrastive learning(DenseCL),which implements self-supervised learning by optimizing a pairwise contrastive(dis)similarity loss at the pixel level between two views of input images.Compared to the supervised ImageNet pre-training and other self-supervised learning methods,our self-supervised DenseCL pretraining demonstrates consistently superior performance when transferring to downstream dense prediction tasks including object detection,semantic segmentation and instance segmentation.Specifically,our approach significantly outperforms the strong MoCo-v2 by 2.0%AP on PASCAL VOC object detection,1.1%AP on COCO object detection,0.9%AP on COCO instance segmentation,3.0%mIoU on PASCAL VOC semantic segmentation and 1.8%mIoU on Cityscapes semantic segmentation.The improvements are up to 3.5%AP and 8.8%mIoU over MoCo-v2,and 6.1%AP and 6.1%mIoU over supervised counterpart with frozen-backbone evaluation protocol.
基金National Natural Science Foundation of China under Grant Nos.61672273 and 61832008Science Foundation for Distinguished Young Scholars of Jiangsu under Grant No.BK20160021+1 种基金Postdoctoral Innovative Talent Support Program of China under Grant Nos.BX20200168,2020M681608General Research Fund of Hong Kong under Grant No.27208720。
文摘Transformers have recently lead to encouraging progress in computer vision.In this work,we present new baselines by improving the original Pyramid Vision Transformer(PVT v1)by adding three designs:(i)a linear complexity attention layer,(ii)an overlapping patch embedding,and(iii)a convolutional feed-forward network.With these modifications,PVT v2 reduces the computational complexity of PVT v1 to linearity and provides significant improvements on fundamental vision tasks such as classification,detection,and segmentation.In particular,PVT v2 achieves comparable or better performance than recent work such as the Swin transformer.We hope this work will facilitate state-ofthe-art transformer research in computer vision.Code is available at https://github.com/whai362/PVT.
