AI Lab authors present research at World’s Leading Conference on Computer Vision

Research paper by AI Lab authors with the “Class Similarity Weighted Knowledge Distillation for Continual Semantic Segmentation” topic is honored to be one of the papers to be presented at the Conference on Computer Vision and Pattern Recognition (CVPR) next June in New Orleans, Louisiana.

The Conference on Computer Vision and Pattern Recognition (CVPR) is the world’s leading conference on Computer Vision, organized annually by the IEEE Computer Society and the Computer Vision Foundation. This is where researchers share the latest research on object recognition, image segmentation, motion estimation, 3D object reconstruction, etc.

The research paper was carried out by a group of authors including Ta The Anh, Tran Thanh Long, Phan Minh Hieu, and Phung Son Lam. The special feature of AI Lab’s work is developing a new method that makes image segmentation with updated data more effective than previous models.

In particular, Deep Learning models clearly show limitations when they are approached with new layers. Continual learning for semantic segmentation (CSS) is an emerging field in computer vision. One problem the authors’ ‘ve noticed in CSS is confusion when it comes to distinguishing old and new classes since they’re both visually similar. To solve this problem, the research team proposes a new CSS framework that applies a class similarity weighted knowledge distillation (CSW-KD) method. CSW-KD will help select the knowledge of the old model, over the old layers – which also appear on the new model.

The benefits of this approach include selective modification of potentially neglected old data and better learning of new classes by associating them with previously seen layers. Extensive tests on the Pascal-VOC 2012 and ADE20k datasets show that this method outperforms other modern methods on standard CSS settings at 7.07% and 8.9%, respectively.

Talking about the next project in project development, Mr. Ta The Anh – representative of the author group shared: “The whole team hopes to be able to develop models for the image segmentation problem on labeled data is missing and geared towards cases that are more relevant to real-world applications”.