The issue lies with the performance of the model.

Hello, I am honored to reproduce your work. However, I encountered an issue during the reproduction process. I tried to use the automatic segmentation code and demo website provided in the official documentation to segment images from the autonomous driving KITTI dataset. I found that the results from the code and the demo website are completely different. The performance of the code seems to be poor. In the code, I set the number of sampling points to be 32*32 and used the VIT-H model. Below is an image showing the segmentation results. I am eagerly awaiting your response.

Start by comparing the automatic segmentation code and demo website with the KITTI dataset, using the stated 32*32 sampling points and ViT-H model settings. Reproduce the differing segmentation results and determine which input or configuration causes the discrepancy; done means the difference is explained and the code and demo produce consistent results or the remaining limitation is documented.

Written by the indexing model from the issue text.