A 3D human pose is estimated from a monocular image using a retrieval-combination approach that exploits the broad capability of example-based approaches and the flexibility of parts-based approaches. Instead of storing and searching for similar full-body examples, we adopt a half-body representation (i.e., either upper-body or lower-body) to reduce a large full-body database into a compact half-body database. The database can create millions of poses by valid half-body combinations. For a given query image, half-body candidates are first retrieved from the database by partial-shape matching. Valid half-body combinations of these candidates are selected based on a learned combination constraint, and then the optimal combination(s) is(are) chosen in a coarse-to-fine evaluation method. We show good experimental results for estimating poses with both synthetic and real images. Our approach has less time and space complexities than example-based approaches and ensures more realistic 3D pose estimates than those of parts-based approaches.