The paper addresses the challenging problem of tracking curves in dense visual clutter, a task that Kalman filters struggle with due to their Gaussian density limitations. The authors propose the CONDENSATION algorithm, which combines factored sampling with learned dynamical models to propagate an entire probability distribution for object position and shape over time. This approach is more robust than Kalman filters, especially in cluttered environments, and can handle agile motion. The algorithm uses 'factored sampling' to represent the distribution of possible interpretations and combines it with a stochastic differential equation model for object motion. The CONDENSATION algorithm is demonstrated to be effective in tracking complex, multi-modal distributions, including rapid motions and jointed objects, outperforming Kalman filters in cluttered scenes. The method is computationally efficient, running in near real-time, and its performance improves with larger sample sizes.The paper addresses the challenging problem of tracking curves in dense visual clutter, a task that Kalman filters struggle with due to their Gaussian density limitations. The authors propose the CONDENSATION algorithm, which combines factored sampling with learned dynamical models to propagate an entire probability distribution for object position and shape over time. This approach is more robust than Kalman filters, especially in cluttered environments, and can handle agile motion. The algorithm uses 'factored sampling' to represent the distribution of possible interpretations and combines it with a stochastic differential equation model for object motion. The CONDENSATION algorithm is demonstrated to be effective in tracking complex, multi-modal distributions, including rapid motions and jointed objects, outperforming Kalman filters in cluttered scenes. The method is computationally efficient, running in near real-time, and its performance improves with larger sample sizes.