Slow inference of the segmentation model

[mengpengfei]

Using the trained image semantic segmentation model, the inference on Nucense is extremely slow – a full run takes about 4 days. What settings can I adjust to speed it up?

[mengpengfei]

Can we generate pseudo-labels on the nuScenes mini dataset using a trained image semantic segmentation model and quickly verify their quality?

[christianll9]

Hi @mengpengfei, we haven't optimized the inference code and used it almost as is from the RoGS repository. For a quick label check you can follow my instructions mentioned here.

[mengpengfei]

I'd like to quickly verify the results, run through the whole pipeline swiftly, and then do some development. Can I change 1.1.6 to perform inference only on the nuScenes mini dataset and then generate pseudo-labels solely based on the nuScenes mini dataset?

[christianll9]

Yes sure, that should also work.

[mengpengfei]

What modifications are needed to use the nuScenes mini dataset?