Jie Hu

I currently serve as a research fellow at the National University of Singapore, where I collaborate with Professor Xinchao Wang. Previously, I worked as an engineer and postdoctoral researcher at Contemporary Amperex Technology Co. Limited, under the guidance of Professor Jun Ni. I earned my Ph.D. degree from Xiamen University, with my research supervised by Professor Rongrong Ji. My research interests lie in the fields of 2D-3D representation learning, scene understanding, and generation.

Email  /  CV  /  Google Scholar  /  Github

PE3R reconstructs 3D scenes using only 2D images and enables semantic understanding through language.

This paper proposes a transformer-based instance segmentation framework, which encodes masks into embeddings to regress them.

This paper proposes a novel framework, PAIS, that aligns the pseudo-labels of unannotated images with varying class and mask quality for semi-supervised instance segmentation, achieving state-of-the-art results on COCO and Cityscapes datasets.

In this paper, we show the complex task of panoptic segmentation can run over 30 FPS with competitive PQ performance, which is achieved by the proposed YOSO with a feature pyramid aggregator and a separable dynamic decoder.

This paper proposes DistilPose that distills heatmap-based human pose estimation models into regression-based ones to speed up the pipeline.

This paper proposes to disentangle deep neural networks via information bottleneck to understand the inner workings of them.

This paper presents a novel method for image-to-image translation that uses a hierarchical tree to re-arrange the image attributes and achieve style disentanglement.

This paper proposes a new approach that separates a representation into two parts with different mutual information constraints.

This paper proposes to break through the barrier of using features across different visual search systems.