Models are generated from objects identified in video. Each model is evaluated based on knowledge of the objects determined from video analysis, and preferred models are identified based on the evaluations. In some examples, each model could be evaluated by tracking a movement of each object in the video by using each model to track the object from which it was generated, evaluating an ability of each model to identify the objects in the video that are similar to the object from which it was generated, and determining an amount of false identifications made by each model of different objects in different video that does not include the object from which it was generated.