Yinsong Xu

Yinsong Xu

Auto-generated profile, based on publicly available information

Yinsong Xu

National Institute of Health Data Science, Peking University, Beijing, China

Research Focus & Metrics

Academic Career Length

8Years (2019-Present)

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

Representative Publications

View All Publications
22 October 2024 in IEEE Transactions on Image Processing
Verified
Journal Article
Closed access
YXYinsong Xu
Yinsong Xu
JTJiaqi Tang
Jiaqi Tang
AMAidong Men
Aidong Men
QCQingchao Chen
Qingchao Chen
Medical image segmentation is a critical task in clinical applications. Recently, the Segment Anything Model (SAM) has demonstrated potential for natural image segmentation. However, the requirement for expert labour to provide prompts, and the domain gap between natural and medical images pose significant obstacles in adapting SAM to medical images. To overcome these challenges, this paper introduces a novel prompt generation method named EviPrompt. The proposed method requires only a single reference image-annotation pair, making it a training-free solution that significantly reduces the need for extensive labelling and computational resources. First, prompts are automatically generated based on the similarity between features of the reference and target images, and evidential learning is introduced to improve reliability. Then, to mitigate the impact of the domain gap, committee voting and inference-guided in-context learning are employed, generating prompts primarily based on human prior knowledge and reducing reliance on extracted semantic information. EviPrompt represents an efficient and robust approach to medical image segmentation. We evaluate it across a broad range of tasks and modalities, confirming its efficacy. The source code is available at https://github.com/SPIresearch/EviPrompt.

1 October 2021 in Knowledge-Based Systems
Verified
Journal Article
Open access
YXYinsong Xu
Yinsong Xu
ZJZhuqing Jiang
Zhuqing Jiang
AMAidong Men
Aidong Men
HWHaiying Wang
Haiying Wang
HLHaiyong Luo
Haiyong Luo

2 May 2024 in IEEE Internet of Things Journal
Verified
Journal Article
Closed access
YXYinsong Xu
Yinsong Xu
QCQingchao Chen
Qingchao Chen
Radar technology has emerged as a pivotal component for various applications within the Internet of Things (IoT). To promote the understanding and integration of radar sensing in developing multimodal applications, we introduce the Radar Can See and Hear (RACER) data set. This data set encompasses synchronized radar sensing, audio, and visual data. Radar, with its capability to detect vocal cord vibrations and lip movements, addresses scenarios where conventional microphone and camera setups may falter, such as through-wall or non-line-of-sight sensing. Specifically, the radar discerns and characterizes human lip and vocal cord movements in the range-Doppler domain. We employ deep neural networks to capture the inherent relationships among radar signatures, audio vocal sound, and visual lip movements during human pronunciations. We evaluate the performances of radar sensing using experiments on speech classification, cross-modality retrieval among audio, video, and radar, and cross-modality distillation from video or audio to radar. We summarize the findings and the limitations of using radar sensing in speech-related multimodal analysis applications. Our codes are available at: https://github.com/SPIresearch/RACER .

23 May 2025 in IEEE Transactions on Pattern Analysis and Machine Intelligence
Verified
Journal Article
Closed access
YXYinsong Xu
Yinsong Xu
AMAidong Men
Aidong Men
YLYang Liu
Yang Liu
Xiahai ZhuangXiahai Zhuang
Xiahai Zhuang
QCQingchao Chen
Qingchao Chen
In deep learning, initializing models with pre-trained weights has become the de facto practice for various downstream tasks. Many unsupervised domain adaptation (UDA) methods typically adopt a backbone pre-trained on ImageNet, and focus on reducing the source-target domain discrepancy. However, the impact of pre-training on adaptation received little attention. In this study, we delve into UDA from the novel perspective of pre-training. We first demonstrate the impact of pre-training by analyzing the dynamic distribution discrepancies between pre-training data domain and the source/ target domain during adaptation. Then, we reveal that the target error also stems from the pre-training in the following two factors: 1) empirically, target error arises from the gradually degenerative pre-trained knowledge during adaptation; 2) theoretically, the error bound depends on difference between the gradient of loss function, i.e., on the target domain and pre-training data domain. To address these two issues, we redefine UDA as a three-domain problem, i.e., source domain, target domain, and pre-training data domain; then we propose a novel framework, named TriDA. We maintain the pre-trained knowledge and improve the error bound by incorporating pre-training data into adaptation for both vanilla UDA and source-free UDA scenarios. For efficiency, we introduce a selection strategy for pre-training data, and offer a solution with synthesized images when pre-training data is unavailable during adaptation. Notably, TriDA is effective even with a small amount of pre-training or synthesized images, and seamlessly complements the two scenario UDA methods, demonstrating state-of-the-art performance across multiple benchmarks. We hope our work provides new insights for better understanding and application of domain adaptation.

24 May 2019
Verified
Conference Paper
Open access
QLQinfeng Li
Qinfeng Li
DLDong Liang
Dong Liang
YXYinsong Xu
Yinsong Xu
NXNaifu Xiao
Naifu Xiao
YLYilin Li
Yilin Li

1 December 2019
Verified
Conference Paper
Closed access
YXYinsong Xu
Yinsong Xu
ZJZhuqing Jiang
Zhuqing Jiang
AMAidong Men
Aidong Men
JPJiangbo Pei
Jiangbo Pei
GJGuodong Ju
Guodong Ju
BYBo Yang
Bo Yang
Part-based network is an effective method to improve performance in person re-identification (re-ID). Most existing methods assume the availability of well-aligned person bounding box images as model input. However, automatic detection in some datasets causes misalignment which negatively affects the performance. In this work, we propose an Attentional Part-based CNN (AP-CNN) model which combines learning partial features and attention selection. First, we partition feature map into several horizontal stripes. Second, we use attention selection in each stripe to align the pedestrian images. Inside, we introduce a free-parameter attention model with skip-layer connection which maximizes the complementary information of different levels without increasing the complexity of network. Results on four datasets validate the competitiveness of AP-CNN over the state-of-the-art achieving Rank-1 accuracy of 94.4% on Market-1501, 87.3% on DukeMTMC-ReID, 73.7% on CUHK03-labeled and 72.6% on CUHK03-detected.