Reproduce “GeoDiffusion: Text-Prompted Geometric Control for Object Detection Data Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Kai Chen, Enze Xie, Zhe Chen, Yibo Wang, Lanqing Hong, Zhenguo Li, Dit-Yan Yeung
GEODIFFUSION introduces a framework that translates geometric conditions into text prompts to guide pre-trained text-to-image diffusion models for object detection data generation. The method improves object detector performance when used for augmentation, particularly for rare classes, and significantly enhances generated image quality and layout consistency.
Key insights
- Complex geometric conditions (like bounding boxes and camera views) can be effectively translated into text prompts and leveraged by pre-trained text-to-image diffusion models for controllable generation.
- This is the first layout-to-image method to conclusively demonstrate that synthetically generated images can consistently improve the training performance of object detectors, especially for annotation-scarce and rare classes.
- Foreground prior re-weighting significantly enhances layout consistency and the generation quality of foreground objects within diffusion models, improving precision for downstream tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.