Performance gap when training on ScanNet and testing on KITTI-360

### Description
Hello, thank you for the excellent work on this project and the accompanying paper.

I am writing to report a significant performance gap that I observed when reproducing cross-dataset evaluation results, and I would like to ask for clarification on whether this behavior is expected.

### Experimental Setup

- Training dataset: /data/agile3d/ScanNet

- Testing dataset: /data/agile3d/KITTI360

- Training procedure:

- Followed the official training pipeline and configuration provided in the repository

- No additional data augmentation or architectural modifications

### Evaluation Results

When evaluating the model trained on ScanNet on the KITTI-360 test set, I obtained the following results:

```
2025-12-28 16:11:57,805 - INFO - Averaged metrics, num_queries: 1.0000,
avg_voxel_iou_1: 0.1941, avg_point_iou_1: 0.1933, mask_bce_1: 0.3533, mask_dice_1: 0.7933,
avg_voxel_iou_2: 0.4215, avg_point_iou_2: 0.4191, mask_bce_2: 0.2348, mask_dice_2: 0.5455,
avg_voxel_iou_3: 0.5110, avg_point_iou_3: 0.5072, mask_bce_3: 0.1901, mask_dice_3: 0.4421,
avg_voxel_iou_4: 0.5601, avg_point_iou_4: 0.5560, mask_bce_4: 0.1687, mask_dice_4: 0.3895,
avg_voxel_iou_5: 0.5877, avg_point_iou_5: 0.5850, mask_bce_5: 0.1509, mask_dice_5: 0.3608,
avg_voxel_iou_6: 0.6106, avg_point_iou_6: 0.6079, mask_bce_6: 0.1483, mask_dice_6: 0.3411,
avg_voxel_iou_7: 0.6298, avg_point_iou_7: 0.6267, mask_bce_7: 0.1414, mask_dice_7: 0.3232,
avg_voxel_iou_8: 0.6434, avg_point_iou_8: 0.6404, mask_bce_8: 0.1364, mask_dice_8: 0.3128,
avg_voxel_iou_9: 0.6434, avg_point_iou_9: 0.6394, mask_bce_9: 0.1379, mask_dice_9: 0.3088,
avg_voxel_iou_10: 0.6499, avg_point_iou_10: 0.6453, mask_bce_10: 0.1355, mask_dice_10: 0.3066
```

### Issue

Compared to the cross-dataset results reported in the paper, these numbers are significantly lower.

I would like to ask:

1. Is this performance degradation expected when directly transferring a model trained on ScanNet to KITTI-360 without domain adaptation?

1. Are there any additional preprocessing steps, dataset-specific configurations, or training tricks (e.g., class remapping, coordinate normalization, voxel resolution adjustments) that are required to reproduce the paper’s results?

1. Were the paper’s cross-dataset results obtained using fine-tuning, mixed-dataset training, or any extra supervision not explicitly mentioned?

### Additional Information

- I verified that the dataset paths and splits are correct.

- The evaluation code and metrics computation follow the official implementation.
Any clarification or guidance would be greatly appreciated.
Thank you very much for your time and for releasing this codebase.