Code release

[Dingrui-Wang]

Amazing work!

When do you plan to release your code? btw, is your code based on mmcv?

Thank you!

[immel-f]

Thanks!

The code is based on M3TR, which itself is based on MapTRv2, so it is based on mmcv and mmdetection3d (albeit fairly old versions).

I plan to release the code in the next month, but definitely before the NeurIPS. The cleanup of the code and making the OSM tag dataset available takes a bit of time, sorry for that.

[Dingrui-Wang]

Is it possible to reproduce the same result without using mmcv and mmdetection3d. These two package are old and not maintained anymore, also hard to deploy training on recent GPUs such as rtx 4090.

[immel-f]

The NLP encoder is based on Huggingface and Sentence Transformers and independently usable. The HD map construction model uses mmcv and mmdetection3d heavily. It would theoretically be possible to rewrite it without these libraries, but very difficult since MapTR uses some custom CUDA ops from mmdetection3d.

However, we used SDTagNet with this setup on RTX A6000 Ada GPUs, which have the same compute capability as RTX 4090, without issues. We even managed to get it running on H100 GPUs but that requires some ugly hacks 😄