Reproduce “Unified Perceptual Parsing for Scene Understanding” 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
Tete Xiao, Yingcheng Liu, Bolei Zhou, Yuning Jiang, Jian Sun
This paper presents Unified Perceptual Parsing (UPP), a comprehensive approach that enables a single system, UPerNet, to simultaneously understand visual concepts across five hierarchical levels: scene, object, part, material, and texture. The method leverages a new unified dataset, Broden+, and demonstrates strong performance on individual tasks while also revealing compositional visual knowledge learned from heterogeneous data.
Key insights
- The UPerNet architecture, combining FPN and PPM, efficiently processes multi-scale features for diverse visual concepts.
- A training strategy involving random data source sampling and selective parameter updates effectively handles heterogeneous annotations from multiple datasets.
- The trained model can discover and articulate hierarchical visual knowledge, such as common scene-object, object-material, and material-texture relationships, demonstrating an understanding of compositional scene structure.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.