Asking about the zeroshot performance on semanticKITTI dataset

[bqm1111]

In your paper, it is reported that the method achieved 7.28% meanIoU on semanticKITTI dataset. How should this result be interpreted in practice? In particular, is this performance considered competitive for the open-vocabulary or zero-shot setting, and what are the main factors that explain the relatively low absolute mIoU? It would also be helpful to understand which baseline methods provide the most meaningful comparison and what this result demonstrates about the method’s strengths and limitations.

[anhquancao]

Hi, and thanks for the question

7.28 mIoU should be read as a zero-shot, open-vocabulary number: the model never trains on SemanticKITTI, never sees its taxonomy, gets no intrinsics/extrinsics or rig info, and isn't fine-tuned.

Note that supervised camera-only SSC (~11–15 mIoU) has all of that, and existing open-vocab occupancy works (POP-3D, VEON, LangOcc, OccNeRF…) drop the labels but still train in-domain with calibration, so the only really comparable setting is cross-dataset transfer, where calibration-dependent models collapse.

The low absolute value can be explained by:

• Geometry caps the metric (25.9 IoU on KITTI which means a voxel must be correctly occupied before its label counts) and metric scale is ambiguous because camera parameters aren't an input.
• Supervision comes from sparse LiDAR, so thin and distant structures are weakly constrained.
• Class names themselves are ambiguous as prompts, and the fine taxonomy is where the loss concentrates (7.28 at 19 classes vs 13.53 at 6 superclasses).
• the distilled semantic features still trail the pretrained teacher and training covers only sequence or surround regimes, leaving spatio-temporal ambiguity at inference.

Hope this helps