Could you share some insights on why the PointBert as the backbone in ULIP-2 works much better than other candidates?

[auniquesun]

Thanks for sharing ULIP and ULIP-2, they are great work!

I noticed that in ULIP-2, the zero-shot accuracy of PointBert is much higher than that of PointNeXt on ModelNet40. But after finetuning, the PointNeXt outperforms the PointBert on ScanObjectNN.

Could you provide some insights or other experiments to explain the discrepance? As we know, PointNeXt is the lateset published point cloud processing backbone, it should work better than PointBert in zero-shot classification in theory.

Thanks.

[Tycho-Xue]

I think pointnext itself is tailored for lightweight applications, so the number of parameters might be an important factor.