A comparison of active-contour models based on blurring and on marginalization

Many different active-contour models have been proposed over the last 20 years, but very few comparisons between them have been carried out. Further, most of these comparisons have been either exclusively theoretical or exclusively experimental. This paper presents a combined theoretical and experimental comparison between two contour models. The models are put into a common theoretical framework and performance comparisons are carried out on a vehicle tracking task in the PETS test sequences. Using a condensation tracker helps to find the few frames where either model fails to provide a good fit to the image. The results show that: (a) neither model has a definitive advantage over the other; and (b) Kalman filtering might actually be more effective than particle filtering for both models.