Reproduce “PSP-HDRI$+$: A Synthetic Dataset Generator for Pre-Training of Human-Centric Computer Vision Models” 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
Salehe Erfanian Ebadi, Saurav Dhakad, Sanjay Vishwakarma, Chunpu Wang, You-Cyuan Jhang, Maciek Chociej, Adam Crespi, Alex Thaman and 1 more
PSP-HDRI+ is an enhanced synthetic data generator leveraging the Unity engine and extensive domain randomization for pre-training human-centric computer vision models. It consistently outperforms ImageNet pre-training and other synthetic datasets in human detection and 2D keypoint localization tasks, particularly improving out-of-distribution generalization and few-shot learning performance.
Key insights
- Task-specific, manipulable synthetic data can be a more effective pre-training source than large, generalist real-world datasets for specialized human-centric computer vision tasks.
- Careful design choices in synthetic data generation, such as matching real-world labeling distributions and incorporating realistic occlusions, significantly boost model performance.
- Focused diversity (e.g., "simple animations") can be more effective than indiscriminate variety in synthetic data for improving zero-shot performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.