PeopleSansPeople: A Synthetic Data Generator for Human-Centric Computer Vision

원문 초록

In recent years, person detection and human pose estimation have made great strides, helped by large-scale labeled datasets. However, these datasets had no guarantees or analysis of human activities, poses, or context diversity. Additionally, privacy, legal, safety, and ethical concerns may limit the ability to collect more human data. An emerging alternative to real-world data that alleviates some of these issues is synthetic data. However, creation of synthetic data generators is incredibly challenging and prevents researchers from exploring their usefulness. Therefore, we release a human-centric synthetic data generator PeopleSansPeople which contains simulation-ready 3D human assets, a parameterized lighting and camera system, and generates 2D and 3D bounding box, instance and semantic segmentation, and COCO pose labels. Using PeopleSansPeople, we performed benchmark synthetic data training using a Detectron2 Keypoint R-CNN variant [1]. We found that pre-training a network using synthetic data and fine-tuning on various sizes of real-world data resulted in a keypoint AP increase of $+38.03$ ($44.43 \pm 0.17$ vs. $6.40$) for few-shot transfer (limited subsets of COCO-person train [2]), and an increase of $+1.47$ ($63.47 \pm 0.19$ vs. $62.00$) for abundant real data regimes, outperforming models trained with the same real data alone. We also found that our models outperformed those pre-trained with ImageNet with a keypoint AP increase of $+22.53$ ($44.43 \pm 0.17$ vs. $21.90$) for few-shot transfer and $+1.07$ ($63.47 \pm 0.19$ vs. $62.40$) for abundant real data regimes. This freely-available data generator should enable a wide range of research into the emerging field of simulation to real transfer learning in the critical area of human-centric computer vision.