Projects

Large Scale Visual Recognition

Vision: to build an engine to recognize everything in Internet images
Large Scale Visual Recognition in the Internet scale applications is commercially important and technically challenging.
The major challenges brought by a large number of tags and a huge number of live data include:
- Cost of human labeling;
- Large variation of object scales;
- Difficulty of collecting samples for long-tail tags
Our research directions include:
- Data-Efficient Learning (Webly-Supervised Learning, Weakly-Supervised Object Detection, etc.),
- Neual Architecture Design (Multi-Scale Fusion, etc.),
- Few-shot Learning (Instance Recognition, Instance Retrieval, etc.).

J. Yang, W. Chen, L. Feng, X. Yan, H. Zheng and W. Zhang.
Webly supervised image classification with metadata: Automatic noisy label correction via visual-semantic graph.
Proc. ACM Multimedia (MM), 2020. (Oral presentation, Acceptance rate: 8.9%) [paper]

Webly supervised learning of convolutional neural networks with metadata.

J. Yang, L. Feng, W. Chen, X. Yan, H. Zheng, P. Luo and W. Zhang.
Webly supervised image classification with self-contained confidence.
Proc. European Conf. on Computer Vision (ECCV), 2020. (Acceptance rate: 27%) [paper] [code]

Samplewise web label correction with model confidence and self label.

H. Zheng, L. Feng, Y. Chen, W. Chen, Z. Huang, Z. Sun and W. Zhang.
First runner-up in WebVision Challenge 2019.
[results] [Presentation]

C. Lin, S. Wang, D. Xu, Y. Lu and W. Zhang.
Object instance mining for weakly supervised object detection.
Proc. AAAI Conference on Artificial Intelligence (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.

Y. Li, Z. Kuang, Y. Chen and W. Zhang.
Data-driven neuron allocation for scale aggregation networks.
Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2019. [paper] [github] [blog (in Chinese)]

Learnable scale-adaptive building blocks (an alternative to Residual Blocks) for deep convolutional neural networks.

X. Wang, S. Zhang, Z. Yu, L. Feng and W. Zhang.
Scale-equalizing pyramid convolution for object detection.
Proc. IEEE Conf. on Computer Vision and Pattern Recognition (CVPR), 2020. [blog (in Chinese)] [code]

Pyramid convolution (a kind of 3D convolution) to extract multiscale features as a better alternative design of Feature Pyramid Networks (FPN).

Y. Gao, Z. Kuang, G. Li, P. Luo, Y. Chen, L. Lin and W. Zhang.
Fashion retrieval via graph reasoning networks on a similarity pyramid.
IEEE Transactions on Pattern Analysis and Machine Intelligence (TPAMI), 2020. [doi]
Z. Kuang, Y. Gao, G. Li, P. Luo, Y. Chen, L. Lin and W. Zhang.
Fashion retrieval via graph reasoning networks on a similarity pyramid.
Proc. International Conf. on Computer Vision (ICCV), 2019. (Oral presentation)

Z. Chen, Z. Kuang, W. Zhang and K.-Y. K. Wong.
Learning local similarity with spatial relations for object retrieval.
Proc. ACM Multimedia (MM), 2019.

Z. Chen, Z. Kuang, K.-Y. K. Wong and W. Zhang.
Aggregated deep feature from activation clusters for particular object retrieval.
Proc. ACM Multimedia (MM) Thematic Workshops, 2017.