Josie (Jiaojiao) Zhao

I am currently a final-year PhD candidate on video understanding under the supervision of Prof. Cees Snoek, in the VIS Lab, University of Amsterdam, The Netherlands.

Email  /  CV  /  Google Scholar  /  Twitter  /  Github

2022/03. Our paper on TubeR for video action detection based on transformer has been accepted as oral in CVPR 2022 .

2021/01. Our paper on Liftpool for image classification and segmentation has been accepted in ICLR 2021.

2020/06. I have joined Amazon Rekognition as a reseacher intern.

2019/10. I have joined Kepler Vision Technologies as a reseacher assistant.

2019/06. Our paper on pixelated semantic colorization has been accepted in IJCV.

2019/03. Our paper on action detection with two-in-one stream has been accepted in CVPR 2019.

2018/08. Our paper on image colorization based on pixel-level semantics has been accepted as oral in BMVC 2018.

2017/12. I have joined in VIS Lab as a PhD at UVA.

2021/03. I attend Google International Women's Day with students.

2019/09. I attend BMVA Symposium on Video Understanding 2019 in London.

2019/06. Our paper on action detection with two-in-one stream is presented at Facebook's AI video summit.

2018/09. I attend Computer Vision Summit 2018 at Google in Zurich.

I'm interested in computer vision and deep learning. During my PhD, I am mainly focusing on video understanding, specifically video action detection, action recognition, video object segmentation. I am also excited about getting inspiration from classical signal/image processing methodologies for addressing some basic problems in deep learning. Besides, I have worked on image colorization, face recognition/detection/aligment, and change detection.

Has developped the first Transformer-based end-to-end action tube detection framework.

By adopting the philosophy of the classical Lifting Scheme from signal processing, we propose LiftPool for bidirectional pooling layers, including LiftDownPool and LiftUpPool.

We present aligned 3D convolution blocks, which collect the valuable information from the locations aligned by the learned offsets rather than the original dislocated positions.

We propose to exploit pixelated object semantics to guide image colorization.

We propose to embed RGB and optical-flow into a single two-in-one stream network with new layers for video action detection.

To address context confusion, we propose to incorporate the pixel-level object semantics to guide the image colorization.

We propose a novel set-to-set (S2S) distance measure to calculate the similarity between two sets with the aim to improve the recognition accuracy for faces with real-world challenges, such as extreme poses or severe illumination conditions.

First work applying deep learning to change detection for synthetic aperture radar images.

We establish a deep neural network using stacked Restricted Boltzmann Machines (RBMs) to analyze the difference images and detect changes between multitemporal synthetic aperture radar (SAR) images.

We propose a novel difference image analysis approach based on deep neural networks for image change detection problems.