This invention, which relates to retrieving an object from a video or a photo where the object matches a hand-drawn sketch, discloses a method for automatically estimating a perceptual bias level with respect to a feature of the sketch. The method allows estimation based on the sketch alone without involving an extra database. In one embodiment, the method comprises using an expectation-maximization tensor voting (EMTV) method to analyze a statistical distribution of the feature. The statistical distribution is analyzed by forming an objective function having the statistical distribution's information parameterized by the perceptual bias level, and then maximizing the objective function according to a set of iterative update rules. In another embodiment, the method for automatically estimating a perceptual bias level is incorporated into a method for retrieving one or more objects from an image or video database where the one or more objects match a hand-drawn sketch.