Something wrong when I want to compute the depth metrics on all frames in each sequence about scenerf

[BeileiCui]

Hi @anhquancao I tried your suggestion to change "squeeze()" to "squeeze(-1)" and the problem is solved! Really thanks for your time!

[anhquancao]

Hi @wangqiwangyuqing, thank you for pointing out the error, let me check it and will get back to you this week.

[anhquancao]

It's weird that you have ~4k frames. For validation, there are only 800 frames since we only use the frame_id % 5 == 0 as in this line of code.
https://github.com/astra-vision/SceneRF/blob/main/scenerf/data/semantic_kitti/kitti_dataset.py#L95

Can you check if the variable self.scans in the following line contains only frames with frame_id % 5 == 0
https://github.com/astra-vision/SceneRF/blob/main/scenerf/data/semantic_kitti/kitti_dataset.py#L165

[BeileiCui]

@anhquancao Hi. I met the same error and my validation only contains frame_id % 5 == 0

looking forward to your reply!

[anhquancao]

Hi @BeileiCui, thanks for your info.
It's really weird, I cloned the code, downloaded the checkpoint and ran the evaluation. However, my evaluation passed item 92/799 and had no error.

Did you following my instruction to setup the dataset?
https://github.com/astra-vision/SceneRF#dataset

Which command did you use?

Can you print the length of weights_at_depth and the shape of the tensor inside?

[anhquancao]

Can you change the line 737 as following in scenerf.py and try again @wangqiwangyuqing @BeileiCui ?
https://github.com/astra-vision/SceneRF/blob/main/scenerf/models/scenerf.py#L737

The error is when there is only 1 ray, the squeeze() function turns the array of 1 item into a scalar which causes the issue in torch.cat. So the solution is to change "squeeze()" to "squeeze(-1)".

[anhquancao]

Great! You are welcome.