Lei Ma

Lei Ma

Auto-generated profile, based on publicly available information

Lei Ma

School of Artificial Intelligence, Beijing Polytechnic, China

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
26 January 2024
Verified
Conference Paper
Open access
GJGuanghao Jin
Guanghao Jin
JWJieying Wang
Jieying Wang
YWYuqing Wang
Yuqing Wang
HDHui Du
Hui Du
LMLei Ma
Lei Ma
QSQingzeng Song
Qingzeng Song
To increase the accuracy of pose estimation on multiple datasets, this paper proposes a new pose estimation method based on the fusion of multiple models. This method consists of two main steps: firstly, a dataset classification by YoloV5 algorithm is performed that is to classify the dataset that may contain the testing samples; secondly, the key point detection is performed by using a deep learning model-based pose estimation that is trained on the predicted dataset. We evaluated the methods on three publicly datasets: Leeds Sports Pose (LSP) dataset, CrowdPose dataset and AnimalPose one. As the experimental results show, our method can achieve the scalability on multiple datasets while ensuring high accuracy.

26 January 2024
Verified
Conference Paper
Open access
GJGuanghao Jin
Guanghao Jin
YWYuqing Wang
Yuqing Wang
JWJieying Wang
Jieying Wang
HDHui Du
Hui Du
LMLei Ma
Lei Ma
QSQingzeng Song
Qingzeng Song
In some deep learning-based applications, we can use the distribution of labels to increase the classification accuracy like the distribution of the labels in the corresponding application. This kind distribution can be obtained by the experience of human being while it is generally imprecise. Using validation set can be a solution to compute precise distribution of labels while this is affected by the application change. In this paper, we propose a novel system, which efficiently utilizes the experience of human being and validation set to improve the performance of classification models. Firstly, our method selects and trains an existing deep learning model on the training set. Then, these models can output the probability of labels on each testing sample. Secondly, our system utilizes the user input and the validation set, which is to get the precise distribution of labels in an application. Finally, the computed distribution of labels is selectively combined with the probability of labels on testing samples to get the final classification results. We selected some existing deep learning models on three public datasets to evaluate the models. As the evaluation results show, our system can improve the performance of classification tasks, which is obviously higher than state-of-art methods.