DITTO is a method for one-shot imitation learning from human demonstrations, enabling robots to replicate tasks based on a single human demonstration. The approach involves two stages: first, extracting the trajectory of the demonstration by segmenting objects and estimating their relative motion. Second, generating a trajectory for a new scene by re-detecting objects and transforming the original trajectory. The method leverages several auxiliary models for segmentation, relative object pose estimation, and grasp prediction. It is evaluated on a variety of tasks, including pick-and-place and articulated object manipulation, and tested on a real robot system. The method is effective in real-world scenarios and the code is publicly available. The paper discusses the challenges of teaching robots through imitation learning, highlighting the difficulties of collecting demonstrations using teleoperation or kinesthetic teaching. DITTO addresses these challenges by focusing on object poses rather than end-effector actions, allowing for more flexible and robust transfer of trajectories. The method is evaluated on both offline and real-world settings, demonstrating its effectiveness in transferring trajectories from human demonstrations to robots. Key contributions include a novel, modular method for one-shot transfer from RGB-D human manipulation demonstrations to robots, experiments validating the method and its ablations, and open-source data and code for evaluation. The method is compared with other approaches in the literature, showing its strengths in handling object-centric trajectories and re-detection. The paper also discusses the challenges of real-world implementation, such as kinematic constraints and sensor noise, and provides insights into the performance of the method on various tasks. Overall, DITTO offers a promising approach for one-shot imitation learning in robotics.DITTO is a method for one-shot imitation learning from human demonstrations, enabling robots to replicate tasks based on a single human demonstration. The approach involves two stages: first, extracting the trajectory of the demonstration by segmenting objects and estimating their relative motion. Second, generating a trajectory for a new scene by re-detecting objects and transforming the original trajectory. The method leverages several auxiliary models for segmentation, relative object pose estimation, and grasp prediction. It is evaluated on a variety of tasks, including pick-and-place and articulated object manipulation, and tested on a real robot system. The method is effective in real-world scenarios and the code is publicly available. The paper discusses the challenges of teaching robots through imitation learning, highlighting the difficulties of collecting demonstrations using teleoperation or kinesthetic teaching. DITTO addresses these challenges by focusing on object poses rather than end-effector actions, allowing for more flexible and robust transfer of trajectories. The method is evaluated on both offline and real-world settings, demonstrating its effectiveness in transferring trajectories from human demonstrations to robots. Key contributions include a novel, modular method for one-shot transfer from RGB-D human manipulation demonstrations to robots, experiments validating the method and its ablations, and open-source data and code for evaluation. The method is compared with other approaches in the literature, showing its strengths in handling object-centric trajectories and re-detection. The paper also discusses the challenges of real-world implementation, such as kinematic constraints and sensor noise, and provides insights into the performance of the method on various tasks. Overall, DITTO offers a promising approach for one-shot imitation learning in robotics.