Recent Publications

Bounding boxes uniquely characterize object detection, where a good detector gives accurate bounding boxes of categories of interest. …
PDF

Analyzing model performance in various unseen environments is a critical research problem in the machine learning community. To study …
PDF

We present an interactive visual framework named InternGPT, or iGPT for short. The framework integrates chatbots that have planning and …
PDF

Abstract — This paper studies the benefit of keypoint estimation on object detection. In particular, we focus on the paradigmatic …
PDF

Experience

 
 
 
 
 

Research Intern, supervised remotely by Wenhai Wang

Shanghai AI Laboratory

Feb 2022 – Present Shanghai, China
  • We find a strong positive correlation between the detection mAP and box stability score. This relationship allows us to predict the accuracy of detectors on various real-world test sets without accessing test ground truths, verified on vehicle detection and pedestrian detection tasks. To our knowledge, we are the first to propose the problem of unsupervised evaluation of object detection. The work is submitted to ICLR 2024.
  • To study the problem of analyzing classification performance in various unseen environments, we introduce CIFAR-10-W, which is a testbed with out-of-distribution test sets that have broad coverage of environmental discrepancies. We conduct extensive benchmarking and show that CIFAR-10-W offers interesting insights inherent to domain generalization and model accuracy prediction tasks. The work is submitted to ICLR 2024.
  • Automatic evaluation of the alignment for image generation. Work in progress.
 
 
 
 
 

Research Intern, supervised remotely by Bo Li and Yiru Wang

Urban Computing Group of Sensetime Co., LTD.

May 2021 – Feb 2022 Beijing, China
  • Build efficient transformer module toolkit for CV, borrowing from efficient MSA in NLP (e.g. Linformer).
  • Study the information redundancy phenomenon existing in Transformer based on ViT, including token redundancy, and head redundancy.
  • From multi-scale and feature selection views, research the pyramid structure networks such as Swin and PVT for dense prediction tasks such as detection and segmentation.
 
 
 
 
 

Research Assistant, supervised remotely by A/Prof. Liang Zheng

Australian National University & Seeing Machines Co., LTD

Sep 2020 – May 2021 Canberra, Australian
Study the impact of keypoint branch on one-stage detectors and two-stage detectors, and improve the CenterNet network to solve multiple lightweight detection tasks. The result achieves the state of the art on COCO-Wholebody and WiderFace dataset, and the work was accepted by FG 2021.
 
 
 
 
 

Research Intern, supervised by Prof. Jiaya Jia

Chinese University of Hong Kong & SmartMore Co., LTD.

Jun 2020 – Sep 2020 Shenzhen, China
Multi-Object Tracking. Familiar with detectron2 framework, added RoI expansion, Deepsort tracking module in the two-stage target detector. I practiced FairMot, MaskTrackRCNN and other models. The model was successfully deployed in the border vehicle detection scenario.
 
 
 
 
 

Research Assistant, supervised by Prof. Xinggang Wang

Media and Communications Laboratory, Huazhong University of Science and Technology

Jun 2019 – Sep 2019 Wuhan, China
Different from these methods that considering bounding box as a whole, we research a novel object bounding box representation using points and links.

Contact