FreeSeg: Unified, Universal and Open-Vocabulary Image Segmentation

Figure 1. We propose FreeSeg, a generic framework to accomplish Unified, Universal and Open-Vocabulary Image Segmentation. (a) FreeSeg optimizes an all-in-one network via one-shot training. (b) FreeSeg employs the same architecture and parameters to handle diverse segmentation tasks seamlessly in the inference procedure. (c) FreeSeg establishes new state-of-the-art performance across diverse segmentation tasks, training datasets and zero-shot generalization.