Projects

ViM: Out-of-distribution with virtual-logit matching. CVPR 2022.
[code] [Project Page] [blog (in Chinese)]

ViM is a simple but effective post-hoc OOD detection method, which combines feature space and logit space. It has been included in several courses such as Intro to ML Safety, Recent Advances in Deep Learning, and open source libraries such as OpenOOD, pytorch-ood, Oodeel.
OpenImage-O: A large-scale real-world OOD dataset for ImageNet-1K, which is human-annotated and is 8.8x the size of existing datasets.

OpenOOD: Benchmarking generalized out-of-distribution detection. NeurIPS 2022 Datasets and Benchmarks Track.
[code]

Consistent-Teacher: Towards reducing inconsistent pseudo-targets in semi-supervised object detection. CVPR 2023 (Highlight).
[code]

Revisiting weak-to-strong consistency in semi-supervised semantic segmentation. CVPR 2023.

Uncertainty estimation via response scaling for pseudo-mask noise mitigation in weakly-supervised semantic segmentation. AAAI 2022.
[code]

Pseudo-mask matters in weakly-supervised semantic segmentation. ICCV 2021.
[code]

Object instance mining for weakly supervised object detection. AAAI 2020.
[github] [blog (in Chinese)]

The first weakly supervised object detection (WSOD) approach to utilize multiple instances of the same object category in a image.

Webly supervised image classification with metadata: Automatic noisy label correction via visual-semantic graph. ACM MM 2020. (Oral)

Webly supervised learning of convolutional neural networks with metadata.

Webly supervised image classification with self-contained confidence. ECCV 2020.
[paper] [code]

Samplewise web label correction with model confidence and self label.

First runner-up in WebVision Challenge 2019.
[results] [Presentation]