📌 Click Title for Details!
This is a pinned post to help you explore more. Click on a post title in Paper Insights for paper details, and click on About Me to learn more about me.
Dual-Domain Triple Contrast for Unsupervised Domain Adaptation
Unsupervised Domain Adaptation (UDA) methods learn from labeled data within source domain and unlabeled data within target domain, and then perform downstream tasks on target domain. Current methods approach UDA task using domain adaptation or self-supervised learning strategies, while our D2TC leverages contrastive learning to integrate both strategies into a unified framework. D2TC is used to perform cross-dataset skeleton-based action recognition for practice evaluation.
FLP-GAN for Text-to-Face Image Sythesis
Utilzing AI models to generate faical image based on textual descriptions is an interest thing in AICG filed. However, directly generating images from text conditional inputs is diffcult because of domain gap. Thus, we proposed the FLP-GAN models, which leverages facial landmark as a semantic bridge to faciliate the generation from text to facial images.
Ensemble Contrastive Learning for Unsupervised Representation Learning
How to learning the meaningful features of unlabeled data in unsupervised manner is a key question in AI models. Thus, we proposed a ensemble contrastive learning framework for unsupervised representation learning, named EnsCLR. Finally, EnsCLR was applied to the unsupervised skeleton-based action representation learning task for practice implementation.
Recurrent Attention Networks for Sequential Weakly Labeled Sensors-based HAR
As mentioned in Attention-based CNN, the weakly labeled data inevitably occurs in the process of sensors-based data collection, which may impact the performance of traditional sensors-based human activity recognition methods. We proposed the recurrent attention networks for sequential weakly labeled sensors-based HAR, taking a further step than the Attention-based CNN method.
Attention-based CNN for Weakly Labeled Sensors-based HAR
The weakly labeled data inevitably occurs in the process of sensors-based data collection, which may impact the performance of traditional sensors-based human activity recognition methods. Thus, we proposed the attention-based CNN, which utilizes an attention mechanism to enhance the CNN backbone model’s generalization ability to the weakly labeled samples.
Older
Newer