Reproduce “ViTA-Seg: Vision Transformer for Amodal Segmentation in Robotics” 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
Donato Caramia, Florian T. Pokorny, Giuseppe Triggiani, Denis Ruffino, David Naso, Paolo Roberto Massenio
A Vision Transformer-based framework, ViTA-Seg, performs class-agnostic amodal segmentation for cluttered industrial environments, achieving superior accuracy and about 12 times faster inference than state-of-the-art methods by using MAE pre-training and a dual-head architecture.
Key insights
- Explicitly predicting the occluded mask via a Dual-Head architecture with cross-task collaboration significantly enhances amodal and occluded segmentation accuracy.
- Leveraging Vision Transformers pre-trained with Masked Autoencoders provides strong generalization capabilities and effective global context reasoning for inferring occluded object parts.
- Synthetic data generated in a physics simulator, combined with a transfer learning strategy, effectively addresses data scarcity and improves performance in industrial bin-picking environments.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.