Cars from rear

On the left we show the shape model. Each part of the model is shown in a different colour. The mean location is indicated by the cross, with the ellipse showing the uncertainty in location. The number by each part is the probability of that part being present. On the right, we show examples of regions which are closest to the mean of the appearance density of each part. This shape model is scale-invariant and was trained on images varying in size from 100-550 pixels. The model seems to pick up on a variety of features on the rear of the car.

Here are some examples of the model fitting new images. For illustration, the images have been scale normalized but in reality vary widely in scale.