Let’s Start Over: Retraining with Selective Samples for Generalized Category Discovery
Zhimao Peng, Enguang Wang, Xialei Liu, Ming-Ming Cheng

• This paper introduces a novel method using Nearest Neighbor Distance-aware Label Consistency sample selection. It creates class-consistent subsets for novel class sample clusters from the current GCD method, acting as “pseudo-labeled sets” to mitigate representation bias.
• It proposes progressive supervised representation learning with selected samples to optimize the trade-off between quantity and purity in each subset.
• The results demonstrate the superiority of the method in achieving improved performance in GCD tasks.

Sharpness-aware Dynamic Anchor Selection for Generalized Category Discovery
Zhimao Peng, Enguang Wang, Fei Yang, Xialei Liu, Ming-Ming Cheng

• This paper introduces a loss sharpness penalty into GCD, which generates more robust pseudo labels for training samples by minimizing the worst-case sharpness.
• It proposes a novel dynamic anchor selection strategy to further improve the performance by assigning hard pseudo-labels to the selected unknown classes anchors.
• Experimental results show a state-of-the-art performance

Predictive sample assignment for semantically coherent out-of-distribution detection
Zhimao Peng, Enguang Wang, Xialei Liu, Ming-Ming Cheng

• This paper analyzes the limitations of existing SCOOD methods based on the clustering-based binary In-distribution Filtering strategy and proposes to use dual-threshold ternary sample assignment to select more reliable auxiliary ID and OOD samples for the SCOOD task.
• It proposes a predictive sample assignment framework for the SCOOD task, which includes a threshold sample assignment mechanism, an ID/OOD concept contrastive representation learning loss, and an additional sample retraining process. Experiments show a state-of-the-art performance.

Early Preparation Pays Off: New Classifier Pre-tuning for Class Incremental Semantic Segmentation
Zhengyuan Xie, Haiquan Lu, Jia-wen Xiao, Enguang Wang, Le Zhang, Xialei Liu

• This paper proposes a new classifier pre-tuning method that learns a transformation from all old classifiers to generate new classifiers with previous knowledge before the formal training process. It optimizes the initialization of transformation matrices, striking a balance between stability and plasticity.
• Results show that the method can be readily integrated into other approaches and significantly enhances their performance.

Bridging Inter-task Gap of Continual Self-supervised Learning with External Data
Haori Lu, Xusheng Cao, Linlan Huang, Enguang Wang, Fei Yang, Xialei Liu

• This paper points out that existing CCSSL methods overlook the issue of inter-task data comparisons, and proposes BGE to incorporate external data into training to address this gap. It proposes a sampling algorithm to sample external data that are relevant to tasks and sufficiently diverse, while also filtering out OOD data.
• Experiments show that BGE can be seamlessly integrated into existing CCSSL methods and consistently yields significant improvement.

Restoring Forgotten Knowledge in Non-Exemplar Class Incremental Learning through Test-Time Semantic Evolution
Haori Lu, Xusheng Cao, Linlan Huang, Enguang Wang, Fei Yang, Xialei Liu

• This paper highlights that the testing stages can help reduce forgetting and proposes a novel test-time semantic evolution method, RoSE, which is based on a tailored auxiliary task designed to restore forgotten knowledge. During test time, it computes the analytical solution rather than the gradient for updating the model, addressing the limitations of online training.
• Extensive experiments show that RoSE outperforms most state-of-the-art NECIL methods across various scenarios.

See more, know more: Richer prior knowledge for novel class discovery
Zhimao Peng, Enguang Wang, Xialei Liu, Ming-Ming Cheng

• This paper proposes a module to generate auxiliary dummy classes to learn richer prior knowledge for NCD tasks, so that the model can generate less noisy pseudo labels in unsupervised clustering. It also introduces a cluster-guided dummy classes synthesis strategy, the number of synthesized dummy classes can be significantly reduced and further improve the NCD performance.
• The proposed method can outperform existing methods by a significant margin