Reproduce “Object-Centric Data Synthesis for Category-level 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
Vikhyat Agarwal, Jiayi Cora Guo, Declan Hoban, Sissi Zhang, Nicholas Moran, Peter Cho, Srilakshmi Pattabiraman, Shantanu Joshi
Researchers at UCLA and Analog Devices developed and evaluated object-centric data synthesis methods, including Diffusion Copy-Paste and 3D Random Placement, to generate training data for category-level object detection. These methods significantly improved detection performance in low-data regimes, particularly with a sequential training approach and when blending diverse synthetic data types.
Key insights
- A sequential training scheme, where models are first finetuned on synthetic data and then on a small fraction of real data, generally outperforms directly mixing synthetic and real data.
- Synthetic data generation, even with simplified approaches, substantially boosts object detection performance in data-scarce regimes, often matching performance achieved with significantly larger quantities of real data.
- Combining diverse synthetic data generation strategies (e.g., mixing different synthetic methods) further enhances model robustness and detection accuracy, compensating for individual method limitations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.