�ݺ�ߣ

This document summarizes research on an object recognition system that uses distinctive intermediate-level features (e.g. automatically extracted 2D boundary fragments) as keys within a local context region. These keys are assembled within a loose global context to identify objects. The system demonstrates good recognition of a variety of 3D shapes, with tests on over 2000 images evaluating performance under increasing clutter, occlusion, and database size. The system represents an improvement over other methods by being robust to occlusion and clutter without requiring whole-object segmentation.