lucabergamini

Hi @danfeiX , That's possible, at least in theory. The only reason it's not implemented is that we need a zarr buffer to write on while simulating, so that's why...

Hi @theDebugger811 > which are instances of `EgoDataset` and can be removed We need to keep a reference to the datasets for visualisation and some future metrics. We could use...

mmm..I've never experienced this issue under linux for the test.zarr (which has around the same size of validate.zarr). > i create another chunk data set without linking to any storage...

> btw, should i better use validate.zarr or test.zarr to evaluate the model? test.zarr is used in the competition, but you don't have GT for that one stored in the...

> i use a local object that is similar to chunk dataset but no link to any storage in zarr_scenes_chop (zarr_utils). After the loop of append (copy). the output_dataset (linked...

Hi @KhushalB, Yeah I've also run the profiler a lot in the past weeks trying to optimise that module. There are two major factors determining the runtime speed currently: -...

Hey @cxz > The real gain would be if it was possible to vectorize across `samples`, or at least pre-compute the world-to-agent-coords transformations ahead of time. Yeah that sadly is...

> agent-centric coordinate system and heading normalization mmm... not sure what you mean by this. The rasters are already normalised on the heading (the agent is always aligned with the...

> do you think it would make sense to rasterize always on the AV coordinate system and then fixing the raster by a 2-D rotation? (I know it's not equivalent...

I'll leave this open so people can add their contributions :) Note that there is also [this interesting discussion](https://www.kaggle.com/c/lyft-motion-prediction-autonomous-vehicles/discussion/180359) on Kaggle about the same "issue"