Very amazing work, but I used TTT3R inference on the nuplan scene, and it seems to have obtained poor results. Is this a problem with my operation? <img width="1014" height="553" alt="Image" src="https://github.com/user-attachments/assets/8ff4924c-a74a-4abb-9a34-c58adff40d92" />