Zhang Le (张乐)

I am a Professor at University of Electronic Science and Technology of China, where I work on Deep Learning. I am always looking for self-motivated bachelor students (mainly final-year project), master students, Ph.D students as well as Postdoctoral Scholars. Please drop me your resume if you are interested. This page may not be maintained and please visit my new homepage. Thanks.

Email  /  Google Scholar  /  LinkedIn

PhD: Nanyang Technological University, Singapore (2012-2016).

MSC: Nanyang Technological University, Singapore (2011-2012).

BEng: University of Electronic Science and Technology of China (2007-2011).

Professor (06/2021-) : School of Information and Communication Engineering, University of Electronic Science and Technology of China (UESTC).

Scientist (09/2018-06/2021) : Institute for Infocomm Research, Agency for Science, Technology and Research (A*STAR).

Postdoc (02/2016-09/2018) : Advanced Digital Sciences Center.

Editorships:
Associate Editor: IET Biometrics.
Guest Editor: "Ensemble Deep Learning" in Pattern Recognition.
Guest Editor: "Deep Learning for Human Activity Recognition" in Neurocomputing.

Organizer:
IJCAI2019: 1st workshop on "Deep Learning for Human Activity Recognition"
IJCAI2020: 2nd workshop on "Deep Learning for Human Activity Recognition"

Reviewer:
Journals: IEEE TPAMI, IJCV, IEEE TIP, IEEE TNNLS, IEEE TCyber, IEEE TPDS, IEEE TEVC, IEEE TIE, IEEE TII, PR, CVIU etc.
Conferences: CVPR, ICCV, ECCV, MM, AAAI, IJCAI, MICCAI, WCCI, etc.

Teaching:
EE2073-Introduction to EEE Design and Project @ NTU

02/2020: We are organizing the second workshop on "Deep Learning for Human Activity Recognition" in IJCAI2020.

02/2020: I am serving as an Associate Editor of IET Biometrics.

01/2019: We organized an worksop on "Deep Learning for Human Activity Recognition" in IJCAI2019. Selected papers (or extensions) could be published on a special issue of "Deep Learning for Human Activity Recognition" at Elsevier Journal, Neurocomputing.

10/2018: We organized an special issue on "Ensemble Deep Learning" in Pattern Recognition.

09/2018: I joined I2R as a scientist.

05/2018: In OMG-Emotion Challenge 2018, our ADSC team's submissions ranked 1st for vision-only arousal/valence prediction and 2nd for overall valence prediction!

I'm interested in machine learning, deep learning, computer vision, image processing and their applications.

We tackle the long-standing problem of imbalanced data distribution by proposing a simple yet effective locality-based learning paradigm. The method is locality-aware in two aspects. First, we introduce a locality-aware data partition (LADP) approach to group the training data into different bins via locality-sensitive hashing. As a result, a more balanced data batch is then constructed by LADP. To further reduce the training bias and enhance the collaboration with LADP, a new data augmentation method called locality-aware data augmentation (LADA) is proposed where the image patches are adaptively augmented based on the loss. We demonstrate the versatility of the proposed method by applying it for crowd counting and adversarial defense.

we propose a novel Two-stream Convolution Augmented Human Activity Transformer (THAT) model to utilize a two-stream structure to capture both time-over-channel and channel-over-time features, and use the multi-scale con-volution augmented transformer to capture range-based patterns.

We first show that the common practice of employing recurrent neural networks (RNNs) to aggregate temporalspatial features for Video-based Person Reid (VPRe-id) may not be optimal. Specifically, with a diagnostic analysis, we show that the recurrent structure may not be effective learn temporal dependencies than what we expected and implicitly yields an orderless representation. Based on this observation, we then present a simple yet surprisingly powerful approach for VPRe-id, where we treat VPRe-id as an efficient orderless ensemble of image based person re-identification problem. More specifically, we divide videos into individual images and re-identify person with ensemble of image based rankers. Under the i.i.d. assumption, we provide an error bound that sheds light upon how could we improve VPRe-id. Our work also presents a promising way to bridge the gap between video and image based person re-identification.

We presents a fast correspondence selection method to effectively separate true correspondences from false ones at high speed by leveraging the motion smoothness constraint.

We provide a simple solution to alleviate the foreground-background imbalance problem in video anomaly detection.

We design a novel and practical Q-learning approach where the converged Q-values have the practical meaning as the actual probabilities of arriving on time so as to improve the accuracy of finding the real optimal path.

We provide a general deep regression framework which mimics ensemble learning with a single model. We demonstrate its effectiveness on several computer vision tasks including corwd counting, age estimation, apparent personality analysis and image super-resolution.

In this paper, we utilize contrast prior, which used to be a dominant cue in none deep learning based SOD approaches, into CNNs-based architecture to enhance the depth information.

This paper evaluates the recently proposed local features, correspondence pruning algorithms, and robust estimators using strictly defined metrics in the context of image matching and fundamental matrix estimation. Comprehensive evaluation results on four large-scale datasets provide insights into which datasets are particularly challenging and which algorithms perform well in which scenarios.

A Bayes filter with neural networks is proposed for the optimal estimation of occupancy based on light sensor data.

An attention based bi-directional long short-term memory for passive human activity recognition using WiFi CSI signals.

An accurate edge detector using richer convolutional features.

A mobile location estimation scheme for realistically mixed LOS/NLOS/LOS-NLOS environments.

An omnidirectional mobile badminton robot, which is composed of mechanical, visual and motion control subsystems.

A new ensemble of classifiers that consists of decision trees and random vector functional link network for multi-class classification

A novel knowledge distilling strategy to improve deep learning with handcrafed features

we introduce a dynamic augmentation technique to train a much deeper CNN for crowd counting. In order to decrease over-fitting caused by limited number of training samples, multitask learning is further employed to learn generalisable representations across similar domains. We also propose to aggregate multi-scale convolutional features extracted from the entire image into a compact single vector representation amenable to efficient and accurate counting by way of "Vector of Locally Aggregated Descriptors" (VLAD).

A robust and parallel RVFL for RSS based indoor positioning.

Novel knowledge distilling strategy to assist visual feature learning in the convolutional neural network for painting style classification.

A technical Demo. A deep Siamese-like network is introduced to predict one's Big-Five personality and arousal-valence emotion from one portrait photo.

A probabilistic deep voxelwise dilated residual network to segment the whole heart from 3D MR images.

We train a fully convolutional network to learn the feature embedding space for each superpixel.

With no extra parameters, we mimic ensemble learning within a single network.

KCC extends KCF to any kernel function and is not limited to circulant structure on training data, thus it is able to predict affine transformations with customized properties.

Ensemble of randomized ConvNets for visual tracking.

Ensemble of KCFs for visual tracking.

A benchmark summarization for my PhD study.

An online SVM for time series signal.

Extend the oblique random forest for regression problems.

Improve the faster criterion in vehicle routing by extending the bi-delta distribution to the binormal distribution.

An incremental oblique random forest.

We addressed the problem of dyadic interaction recognition using multi-modal data by combining all three modalities and uses both person centric features (via proxemic descriptors from 3D joints) and holistic features (via FCNN based color and depth features).

This paper reviews traditional emsemble learning as well as state-of-the-art deep ensemble methods and thus can serve as an extensive summary for practitioners and beginner.

Benchmark evaluation of RVFL.

Oblique Random Forest by fast iterative clustering.

ConvNet for visual tracking.

Source code