Spatiotemporal boundary formation (SBF) refers to perception of continuous contours, shape, and global motion from sequential transformations of widely separated surface elements. How such minimal information in SBF can produce whole forms and the nature of the computational processes involved remain mysterious. Formally, it has been shown that orientations and motion directions of local edge fragments can be recovered from small sets of element changes (Shipley & Kellman, (1997). Vision Research, 37, 1281–1293). Little experimental work has examined SBF in simple situations, however, and no model has been able to predict human SBF performance. We measured orientation discrimination thresholds in simple SBF displays for thin, oriented bars as a function of element density, number of element transformations, and frame duration. Thresholds decreased with increasing density and number of transformations, and increased with frame duration. An ideal observer model implemented to give trial-by-trial responses in the same orientation discrimination task exceeded human performance. In a second group of experiments, we measured human precision in detecting inputs to the model (spatial, temporal, and angular inter-element separation). A model that modified the ideal observer by added encoding imprecision for these parameters, directly obtained from Exp. 2, and that included two integration constraints obtained from previous research, closely fit human SBF data with no additional free parameters. These results provide the first empirical support for an early stage in shape formation in SBF based on the recovery of local edge fragments from spatiotemporally sparse element transformation events.