We present a robust 3-D parts-based (PB) tracking system designed to follow the upper limb of stroke survivors during desktop activities. This system fits a probabilistic model of the arm to sequences of images taken from multiple angles. The arm model defines shapes and colors of limbs and limb configurations that are more or less likely. We demonstrate that the system is 1) robust to cluttered scenes and temporary occlusions, 2) accurate relative to a commercial motion capture device, and 3) capable of capturing kinematics that correlate with concurrent measures of post-stroke limb function. To evaluate the PB system, the functional motion of seven stroke survivors was measured concurrently with the PB system and a commercial motion capture system. In addition, functional motion was assessed by an expert using the Fugl-Meyer Assessment (FMA) and related to recorded kinematics. Standard deviation of differences in measured elbow angles between systems was 5.7^; deviation in hand velocity estimates was 2.6 cm/s. Several statistics, moreover, correlated strongly with FMA scores. Standard deviation in shoulder velocity had a significant correlation coefficient with FMA score below -0.75 when measured with all systems.