MSVP

**[ICCV 2025]**Official PyTorch implementation of "Incremental Few-Shot Semantic Segmentation via Multi-Level Switchable Visual Prompts"

Please visit our Official Website if you are in Chinese Mainland.

[Paper] | [Arxiv] | [Supplementary]

:fire: Highlight

• An innovative VLM-based and prompt-based IFSS framework:We propose the first prompt-based IFSS framework, which introduces textual semantics and visual prompts to encode foreground and background classes separately, enabling incremental semantic segmentation.
• Multi-level switchable visual prompts: We propose multi-level switchable visual prompts that customizes multi-granular knowledge tailored to input images, enhancing the model's ability to learn novel classes while maintaining knowledge of old classes.
• A new SOTA performance: Extensive experiments demonstrate the effectiveness of the proposed method. Under the 1-shot condition, it achieves 49.1% mIoU-N on VOC and 25.6% mIoU-N on COCO, setting a new SOTA performance.

:crown: Overview

Images are inputs into the query function to obtain global query features $q_g$ and pixel-wise query features $q_c$. Stage-specific prompts $p_s$ are generated by an attention-like integration way through $q_g$. Region-unique prompts $p_l$ are generated by nearest neighbor matching through clustered $q_c$. Image tokens, concatenated with these selected prompts, are input into pre-trained models to produce predictions. At incremental training stage, the model extends by fine-tuning novelly-added stage-specific prompts and region-unique prompts.

:art: Code

Coming soon.