Is sonata encoder satisfy transform-invariance?

[zlynpu]

I found that the pre-trained Sonata encoder produces different outputs depending on whether center-shift is applied. How can it be adapted for downstream tasks like matching and registration, where center-shift is usually not applicable? By the way, since lidar point cloud often don't have color, is there any solution for producing transform-invariance features without color?

[Gofinge]

Hey, wish the following responses are helpful:

produces different outputs depending on whether center shift is applied

Yeah, in our experiment, a center shift is needed as we center shift all the point clouds during training. Maybe add some random shift during training will make it more robust regarding shift.

By the way, since lidar point cloud often don't have color, is there any solution for producing transform-invariance features without color?

Just pad color with [0, 0, 0]. In our upcoming research, we will add random drop color and random drop normal during training, make the model adaptive to different inputs, and will enable joint training across object, indoor and outdoor to shape one pre-trained model.

[zlynpu]

Thanks for your explanation！I'll try to pad color with [0,0,0] for color missing situation. And for center shift, maybe just changing the pose of point cloud can solve this problem. Anyway, thanks again for your reply and good work on Sonata.