The recent developments in the research area of open-vocabulary segmentation and continual learning have shown significant advancements, particularly in enhancing the adaptability and robustness of models across diverse domains and tasks. Key innovations include the integration of vision-language models with unsupervised domain adaptation, the development of adaptive prompting strategies for continual learning, and the introduction of novel frameworks that leverage pre-trained models for efficient and effective segmentation. These advancements are addressing critical challenges such as catastrophic forgetting, domain shift, and the need for fine-grained semantic understanding. Notably, methods like Mask-Adapter and DenseVLM are pioneering in their approach to refining mask-based classification and decoupling alignment in dense prediction tasks, respectively. Additionally, continual learning strategies such as SAMCL and CoSAM are demonstrating superior performance in dynamic and streaming data scenarios, ensuring models can adapt without significant loss of prior knowledge. These developments collectively push the boundaries of what is possible in open-vocabulary segmentation and continual learning, offering promising solutions for real-world applications.
Advancing Open-Vocabulary Segmentation and Continual Learning
Sources
Mask-Adapter: The Devil is in the Masks for Open-Vocabulary Segmentation
Unsupervised Segmentation by Diffusing, Walking and Cutting
Prompt Transfer for Dual-Aspect Cross Domain Cognitive Diagnosis
SAMCL: Empowering SAM to Continually Learn from Dynamic Domains
ACQ: A Unified Framework for Automated Programmatic Creativity in Online Advertising
Category-Adaptive Cross-Modal Semantic Refinement and Transfer for Open-Vocabulary Multi-Label Recognition
DenseVLM: A Retrieval and Decoupled Alignment Framework for Open-Vocabulary Dense Prediction
Continual Learning for Segment Anything Model Adaptation
Active Learning with Context Sampling and One-vs-Rest Entropy for Semantic Segmentation
Class Balance Matters to Active Class-Incremental Learning
Knowledge Transfer and Domain Adaptation for Fine-Grained Remote Sensing Image Segmentation
EGEAN: An Exposure-Guided Embedding Alignment Network for Post-Click Conversion Estimation
Crack-EdgeSAM Self-Prompting Crack Segmentation System for Edge Devices
Filling Memory Gaps: Enhancing Continual Semantic Parsing via SQL Syntax Variance-Guided LLMs without Real Data Replay
Adaptive$^2$: Adaptive Domain Mining for Fine-grained Domain Adaptation Modeling
Adaptive Prompting for Continual Relation Extraction: A Within-Task Variance Perspective
CAPrompt: Cyclic Prompt Aggregation for Pre-Trained Model Based Class Incremental Learning
VLMs meet UDA: Boosting Transferability of Open Vocabulary Segmentation with Unsupervised Domain Adaptation
Dynamic Prompt Allocation and Tuning for Continual Test-Time Adaptation
Towards Open-Vocabulary Video Semantic Segmentation
MOS: Model Surgery for Pre-Trained Model-Based Class-Incremental Learning