FlowOVD: Learning Generative Latent Flows for Zero-shot Open-vocabulary Detection
Researchers have introduced FlowOVD, a novel approach to open-vocabulary object detection that reframes the problem from a discriminative to a generative one. This method utilizes a continuous transport process in latent space, transforming static queries into text-guided ones through rectified flow. FlowOVD demonstrates improved performance, particularly on challenging datasets like LVIS, by enabling more expressive semantic alignment without additional training data. AI
IMPACT Introduces a novel generative approach to object detection, potentially improving generalization on long-tailed datasets.