Published Papers

Traditional instance semantic segmentation, based on frameworks like Detectron2, is restricted by a "closed vocabulary" derived from its training data (e.g., COCO), limiting its ability to recognize objects from unseen categories. To overcome this limitation, we present TALOS, a modular and flexible method for open-vocabulary instance semantic segmentation. TALOS executes a sequence of three stages: Tagging (extraction of semantic labels of present object classes), Location (bounding box localization for each instance via visual grounding based on the extracted labels), and Segmentation (generation of accurate pixel masks in a category-agnostic manner). Modularity allows integrating diverse state-of- the-art technologies. Qualitative evaluations demonstrate that TALOS correctly identifies and segments objects from categories beyond COCO, outperforming Detectron2 in semantic richness and mask quality, especially in complex scenes.