2018/11/14 by Rawal Khirodkar, Khirodkar, Rawal, Donghyun Yoo +3 · 1 citation
Computer Science · #Advanced Neural Network Applications #Domain Adaptation and Few-Shot Learning #Multimodal Machine Learning Applications
paper · pdf · doi:10.48550/arxiv.1811.05939
We address the issue of domain gap when making use of synthetic data to train\na scene-specific object detector and pose estimator. While previous works have\nshown that the constraints of learning a scene-specific model can be leveraged\nto create geometrically and photometrically consistent synthetic data, care\nmust be taken to design synthetic content which is as close as possible to the\nreal-world data distribution. In this work, we propose to solve domain gap\nthrough the use of appearance randomization to generate a wide range of\nsynthetic objects to span the space of realistic images for training. An\nablation study of our results is presented to delineate the individual\ncontribution of different components in the randomization process. We evaluate\nour method on VIRAT, UA-DETRAC, EPFL-Car datasets, where we demonstrate that\nusing scene specific domain randomized synthetic data is better than\nfine-tuning off-the-shelf models on limited real data.\n