Label-Free Synthetic Pretraining of Object Detectors
This addresses the problem of reducing annotation costs and computational resources for object detection pretraining, though it is an incremental improvement over existing synthetic data methods.
The authors tackled the problem of pretraining object detectors without semantic labels by proposing SOLID, which uses synthetic images from unlabeled 3D models and an instance detection task. The result showed that pretraining on rendered images achieved performance competitive with real images on COCO while using significantly less computing resources.
We propose a new approach, Synthetic Optimized Layout with Instance Detection (SOLID), to pretrain object detectors with synthetic images. Our "SOLID" approach consists of two main components: (1) generating synthetic images using a collection of unlabelled 3D models with optimized scene arrangement; (2) pretraining an object detector on "instance detection" task - given a query image depicting an object, detecting all instances of the exact same object in a target image. Our approach does not need any semantic labels for pretraining and allows the use of arbitrary, diverse 3D models. Experiments on COCO show that with optimized data generation and a proper pretraining task, synthetic data can be highly effective data for pretraining object detectors. In particular, pretraining on rendered images achieves performance competitive with pretraining on real images while using significantly less computing resources. Code is available at https://github.com/princeton-vl/SOLID.