Why are "extrinsics" presented in the train Method?

[eotich32]

As mentioned in the paper, the NoPoSplat is a pose free 3dgs, but why extrinsics are needed in training_step in model_wrapper.py?

def training_step(self, batch, batch_idx):
...
gaussians = self.encoder(batch["context"], self.global_step, visualization_dump=visualization_dump)
output = self.decoder.forward(
gaussians,
batch["target"]["extrinsics"], # here, it seems that we need extrinsics from the training dataset
batch["target"]["intrinsics"],
batch["target"]["near"],
batch["target"]["far"],
(h, w),
depth_mode=self.train_cfg.depth_mode,
)

these confused me

[eotich32]

I get it now.
The proposed approach needs GT extrinsics to train a NN model on large dataset.
After trained, the model predicts gaussians without extrinsics for any given scenes.
This approach really differs from previous 3DGSs.