MGD-GAN: Text-to-Pedestrian Generation Through Multi-grained Discrimination

In this paper, we investigate the problem of text-to-pedestrian synthesis, which has many potential applications in art, design, and video surveillance. Existing methods for text-to-bird/flower synthesis are still far from solving fine-grained image generation problem, due to complex structure heterogeneous appearance that pedestrians naturally take on. To end, propose Multi-Grained Discrimination enhanced Generative Adversarial Network, capitalizes a human-part-based Discriminator (HPD) self-cross-attended (SCA) global order capture coherence body structure. A fined-grained word-level attention mechanism is employed HPD module enforce diversified vivid details. addition, two pedestrian metrics, named Pose Score Variance, devised evaluate quality diversity. We conduct extensive experiments ablation studies on caption-annotated dataset, CUHK Person Description Dataset. The substantial improvement over various metrics demonstrates efficacy MGD-GAN scenario.