A5100335202- Domain Adaptation and Few-Shot Learning
- Advanced Neural Network Applications
- Multimodal Machine Learning Applications
- Remote-Sensing Image Classification
- Advanced Image and Video Retrieval Techniques
- Adversarial Robustness in Machine Learning
- Generative Adversarial Networks and Image Synthesis
- Video Surveillance and Tracking Methods
- Geophysical Methods and Applications
- Anomaly Detection Techniques and Applications
- Machine Learning and Data Classification
- Human Pose and Action Recognition
- Data Management and Algorithms
- Digital Media Forensic Detection
- COVID-19 diagnosis using AI
- Machine Learning and ELM
- Advanced SAR Imaging Techniques
- Remote Sensing and LiDAR Applications
- Neural Networks and Applications
- Synthetic Aperture Radar (SAR) Applications and Techniques
- Machine Learning and Algorithms
- Topological and Geometric Data Analysis
- Semantic Web and Ontologies
- Robotics and Sensor-Based Localization
- Face and Expression Recognition
Beijing Academy of Artificial Intelligence
2024-2025
Shanghai Artificial Intelligence Laboratory
2023-2025
Fudan University
2007-2024
Didi Chuxing (China)
2023
Shandong University
2023
Zhejiang Sci-Tech University
2023
Civil Aviation University of China
2023
Zhongshan Hospital
2021-2022
Microsoft Research Asia (China)
2020-2021
Donghua University
2021
Self-training is a competitive approach in domain adaptive segmentation, which trains the network with pseudo labels on target domain. However inevitably, are noisy and features dispersed due to discrepancy between source domains. In this paper, we rely representative prototypes, feature centroids of classes, address two issues for unsupervised adaptation. particular, take one step further exploit distances from prototypes that provide richer information than mere prototypes. Specifically,...
Cholangiocarcinoma was a highly malignant liver cancer with poor prognosis, and immune infiltration status considered an important factor in response to immunotherapy. In this investigation, we tried locate related genes of cholangiocarcinoma through combination bulk-sequencing single-cell sequencing technology. Single sample gene set enrichment analysis used annotate datasets TCGA CHOL, GSE32225, GSE26566. Differentially expressed between high- low-infiltrated groups dataset were yielded...
Although domain adaptation has been extensively studied in natural image-based segmentation tasks, the research on cross-domain for very-high-resolution (VHR) remote sensing images (RSIs) still remains underexplored. The VHR RSI-based mainly faces two critical challenges: 1) large area land covers with many diverse object categories bring severe local patch-level data distribution deviations, thus yielding different difficulties patches and 2) sensor types or dynamically changing modes cause...
Global channel pruning (GCP) aims to remove a subset of channels (filters) across different layers from deep model without hurting the performance. Previous works focus on either single task or simply adapting it multitask scenario, and still face following problems when handling pruning: 1) Due mismatch, well-pruned backbone for classification focuses preserving filters that can extract category-sensitive information, causing may be useful other tasks pruned during stage; 2) For...
Few-shot fine-grained learning aims to classify a query image into one of set support categories with differences. Although different objects' local differences via Deep Neural Networks has achieved success, how exploit the query-support cross-image object semantic relations in Transformer-based architecture remains under-explored few-shot scenario. In this work, we propose double-helix model, namely HelixFormer, achieve relation mining bidirectional and symmetrical manner. The HelixFormer...
Simplicity is the ultimate sophistication. Differentiable Architecture Search (DARTS) has now become one of mainstream paradigms neural architecture search. However, it largely suffers from well-known performance collapse issue due to aggregation skip connections. It thought have overly benefited residual structure which accelerates information flow. To weaken this impact, we propose inject unbiased random noise impede We name novel approach NoisyDARTS. In effect, a network optimizer should...
The goal of few-shot fine-grained image classification is to recognize rarely seen objects in the query set, given only a few samples this class support set. Previous works focus on learning discriminative features from limited number training for distinguishing various classes, but ignore one important fact that spatial alignment semantic between with arbitrary changes and image, also critical computing similarity each support-query pair. In work, we propose an object-aware long-short-range...
<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup> Deepfakes raised serious concerns on the authenticity of visual contents. Prior works revealed possibility to disrupt deepfakes by adding adversarial perturbations source data, but we argue that threat has not been eliminated yet. This paper presents MagDR, a mask-guided detection and reconstruction pipeline for defending from attacks. MagDR starts with module defines few criteria judge...
Semi-supervised few-shot learning aims to improve the model generalization ability by means of both limited labeled data and widely-available unlabeled data. Previous works attempt relations between extra data, performing a label propagation or pseudo-labeling process using an episodic training strategy. However, feature distribution represented pseudo-labeled itself is coarse-grained, meaning that there might be large gap real query To this end, we propose sample-centric generation (SFG)...
Autophagy in tumor was also found to influence immune microenvironment. The relation between autophagy and cancer intrinsic PD1 PD-L1 expression not clear.With data from TCGA GTEx databases, mRNA levels of autophagy-related genes were compared samples normal tissues, which correlated with survival status. Expression associated clinical traits datasets GSE14520 ICGC LIRI. Single sample gene set enrichment analysis (ssGSEA) used calculate scores samples, using signatures MSigDB database....
Abstract Deep neural networks (DNNs) have achieved great success in many object detection tasks. However, such DNNS-based large models are generally computationally expensive and memory intensive. It is difficult to deploy them devices with low resources or scenarios high real-time requirements, which greatly limits their application promotion. In recent years, researchers focused on compressing without significantly degrading performance, made progress. Therefore, this paper presents a...
Although remote sensing (RS) data with multiple modalities can be used to significantly improve the accuracy of semantic segmentation in RS data, how effectively extract multimodal information through feature fusion remains a challenging task. Specifically, existing methods for still face two major challenges: 1) Due diverse imaging mechanisms boundaries same foreground may vary across different modalities, leading inclusion unwanted background semantics fused features; 2) from exhibit...
Deep learning has emerged as a transformative approach for solving complex pattern recognition and object detection challenges. This paper focuses on the application of novel framework based RT-DETR model analyzing intricate image data, particularly in areas such diabetic retinopathy detection. Diabetic retinopathy, leading cause vision loss globally, requires accurate efficient analysis to identify early-stage lesions. The proposed model, built Transformer-based architecture, excels at...
Multi-task dense prediction aims at handling multiple pixel-wise tasks within a unified network simultaneously for visual scene understanding. However, cross-task feature interactions of current methods are still suffering from incomplete levels representations, less discriminative semantics in participants, and inefficient pair-wise task interaction processes. To tackle these under-explored issues, we propose novel BridgeNet framework, which extracts comprehensive intermediate Bridge...
<?Pub Dtl=""?> Compared with general videos, movies and TV shows attract a significantly larger portion of people across time contain very rich interesting narrative patterns shots scenes. In this paper, we aim to recover the inherent structure scenes in such video narratives. The obtained could be useful for subsequent analysis tasks as tracking objects cuts, action retrieval, well enriching user browsing editing interfaces. Recent research on problem has mainly focused combining multiple...
Unsupervised domain adaptive object detection aims to adapt a well-trained detector from its original source with rich labeled data new target unlabeled data. Previous works focus on improving the adaptability of region-based detectors, <italic xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">e.g.</i> , Faster-RCNN, through matching cross-domain instance-level features that are explicitly extracted region proposal network (RPN). However, this is...
As a fundamental task of 3D perception, point cloud recognition has shown significant progress in recent years. However, existing methods still face challenges when dealing with geometry differences, resulting performance degradation distribution gap exists between the training and testing data, also known as domain generalization. In this work, we focus on problem propose general framework, named Push-and-Pull, aimed at effectively improving generalization ability models unseen target...
Few-shot object detection aims to localize and recognize potential objects of interest only by using a few annotated data, it is beneficial for remote sensing images (RSIs) based applications such as urban monitoring. Previous RSIs-based few-shot works often try convert the support from class-agnostic features class-specific vectors, then perform feature attention operations on query image be tested. However, methods still face two critical challenges: 1) They ignore spatial similarity...