Did you use the temporal multi-frame inputs when finetuning?

[SeaBird-Go]

Hi, thanks for sharing this wonderful work. Since you use the multi-frame multi-view inputs during pretraining stage, I want to know whether did you still use the temporal multi-frame inputs during fine-tune stage?

If you did not use the temporal multi-frame inputs in the downstream tasks, did it mean you discard the voxel decoder in the finetune stage and only load the pre-trained voxel encoder?

[Doctor-James]

Thank you for your interest in our work. Whether we used temporal multi-frame inputs during fine-tuning depended on whether the methods we were comparing against did so. You can roughly understand it as us providing a pre-trained backbone (such as ResNet50), and during fine-tuning, we adopted exactly the same training strategy as the baseline.