YOLOE

Object detection and segmentation are often constrained by predefined categories or heavy open-set methods. YOLOE consolidates detection and segmentation for text, visual, or no prompts in one efficient model. It re-parameterizes textual embeddings, employs a semantic-activated visual prompt encoder, and leverages a built-in vocabulary for prompt-free detection. Extensive tests show real-time performance, strong zero-shot transferability, and lower training cost. On LVIS, YOLOE outperforms YOLO-Worldv2 with 3× less training cost and faster inference. On COCO, YOLOE exceeds closed-set YOLOv8 with nearly 4× fewer training hours.

Here is an exmple of the model used to identify a variety of objects: