In bootstrap initialization for tracking, we exploit a weak prior model used to track a target to learn a stronger model, without manual intervention. We define a general formulation of this problem and present a simple taxonomy of such tasks. The formulation is instantiated with algorithms for bootstrap initialization in two domains: In one, the goal is tracking the position of a face at a desktop; we learn color models of faces, using weak knowledge about the shape and movement of faces in video. In the other task, we seek coarse estimates of head orientation; we learn a person-specific ellipsoidal texture model for heads, given a generic model. For both tasks, we use nonparametric models of surface texture. Experimental results verify that bootstrap initialization is feasible in both domains. We find that (1)independence assumptions in the learning process can be violated to a significant degree, if enough data is taken; (2)there are both domain-independent and domain-specific means to mitigate learning bias; and (3)repeated bootstrapping does not necessarily result in increasingly better models.