TAG: Guidance-free Open-Vocabulary Semantic Segmentation

TL;DR

TAG addresses open-vocabulary segmentation without training, dense annotations, or user-provided text queries by combining per-pixel representations from DINOv2 with CLIP-derived patch embeddings and retrieving open-vocabulary categories from an external caption database. It introduces a three-stage pipeline: segment candidate generation with DINOv2, representative segment embeddings with CLIP, and segment category retrieval from text databases via cosine similarity, enabling management of unseen classes and proper nouns. The method achieves state-of-the-art results on PascalVOC, PascalContext, and ADE20K, notably +$15.3$ $mIoU$ on PascalVOC compared to previous zero-guidance methods, and demonstrates strong generalization on web-crawled images. The approach is flexible and scalable as it can extend to new concepts by updating the external database, without retraining. Code and data will be released.

Abstract

Semantic segmentation is a crucial task in computer vision, where each pixel in an image is classified into a category. However, traditional methods face significant challenges, including the need for pixel-level annotations and extensive training. Furthermore, because supervised learning uses a limited set of predefined categories, models typically struggle with rare classes and cannot recognize new ones. Unsupervised and open-vocabulary segmentation, proposed to tackle these issues, faces challenges, including the inability to assign specific class labels to clusters and the necessity of user-provided text queries for guidance. In this context, we propose a novel approach, TAG which achieves Training, Annotation, and Guidance-free open-vocabulary semantic segmentation. TAG utilizes pre-trained models such as CLIP and DINO to segment images into meaningful categories without additional training or dense annotations. It retrieves class labels from an external database, providing flexibility to adapt to new scenarios. Our TAG achieves state-of-the-art results on PascalVOC, PascalContext and ADE20K for open-vocabulary segmentation without given class names, i.e. improvement of +15.3 mIoU on PascalVOC. All code and data will be released at https://github.com/Valkyrja3607/TAG.

Figures (7)

• Figure 1: Guidance-free Open-Vocabulary Semantic Segmentation. Our TAG can segment an image into meaningful segments without training, annotation, or guidance. It successfully segments structures such as the Leaning Tower of Pisa and the Colosseum. Unlike traditional open-vocabulary semantic segmentation methods, TAG can segment and categorize without text-guidance.
• Figure 2: High-level overview of our TAG architecture. Our TAG can partition images into semantic segments and label each segment with open-vocabulary categories. First, TAG identifies segment candidates using per-pixel features obtained from DINOv2 dinov2. Then, it acquires representative segment embeddings for segment candidates using per-pixel features from a ViT pre-trained with CLIP clip. Finally, the categories are assigned to each candidate segment by retrieving the closest matching sentence from an external database. Note that the input is only the image, with no need to input category candidates as guidance.
• Figure 3: Overview of the flow for each segment. Each segment independently retrieves for category candidates and assigns a category.
• Figure 4: Comparison results with CLIP base open-vocabulary segmentation methods on PascalVOC voc Note that MaskCLIP maskclip and GEM gem uses text guidance while our TAG does not use.
• Figure 5: Qualitative results. We compare GroupViT groupvit, ZeroSeg zeroseg, and our TAG on images containing general objects from PascalContext pc59. This figure indicates that TAG can segment and label correctly.