Reproduce “GenDet: Painting Colored Bounding Boxes on Images via Diffusion Model for Object Detection” 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
Chen Min, Chengyang Li, Fanjie Kong, Qi Zhu, Dawei Zhao, Liang Xiao
GenDet reframes object detection as an image generation task, where a fine-tuned Stable Diffusion model directly paints colored bounding boxes onto input images. With a learning-based post-processing head, the framework achieved an Average Precision of 46.4 on the COCO 2017 dataset, outperforming several state-of-the-art discriminative object detectors.
Key insights
- Large-scale pre-trained diffusion models possess rich implicit knowledge of object structure and spatial relationships that can be explicitly utilized for object detection.
- Reframing object detection as an image-conditioned annotation generation task allows a single generative model to simultaneously perform image synthesis and object perception.
- Integrating a gradient loss and an optimal bounding box scaling ratio during training is crucial for preserving fine-grained geometric structures and improving detection accuracy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.