• 강사 : 김태균 교수 (KAIST 전산학부) 
• 일시 : 2020. 9. 24 (목) 17:00~18:30

Data augmentation has been a long dream for vision and relevant tasks. Generative deep neural network has emerged, however its inherent limitation is not able to generate new data out of data distribution the model has seen. Consequently, the generated data somehow by interpolation does not help improve accuracy of existing models, which are already trained by a large dataset and with good generalisation. In this talk, we present several ideas to improve the quality and diversity of data generation, which in turn successfully boosts up accuracies of visual recognition. The ideas are based on generative adversarial learning, synthetic 3D models, attribute-relational conditions, use of a Physics simulator, and domain transfer, and they will be demonstrated for 3D hand pose, 6D object pose, and face image generation.

* 강연자료