vision3d
I'm working on vision3d (https://github.com/yeetypete/vision3d) a 3D extension to torchvision. vision3d aims to elegantly extend the well-established torchvision API to 3D object detection tasks. My colleagues and I were not happy with the state of 3D object detection frameworks. Things like mmdetection3d, OpenPCDet, etc. try to do way too much than what's needed for a good 3D ML library. They even came up with their own version of pip named "mim"! vision3d focuses on just doing the things that you shouldn't have to author yourself when training a 3D ML model: semantic tensor types (using torchvision's TVTensor abstraction) defining point clouds, images, bounding boxes, etc., transforms, dataloaders, 3D object detection metrics and some (optional) visualization utilities. We also aim for excellent interoperability with torchvision. Almost all torchvision image transforms can be composed together with vision3d's. In vision3d when you compose transforms together we automatically check to ensure the operation does not break the geometric consistency of your 3D scene. See here for a nice interactive example: https://vision3d.dev/auto_examples/transforms/plot_transform... Under the hood vision3d also does a lot of cool stuff to make consuming it as a dependency as easy as possible! For custom CUDA and C++ extensions we use the Torch Stable ABI meaning our single published wheel should work across any PyTorch >= 2.10. We also statically link to the CUDA runtime which allows our wheel to work out of the box on any host which supports CUDA >= 12.8! Currently we are working on typing our API with Tensor shapes with Pyrefly (https://pyrefly.org/en/docs/tensor-shapes/). We aim to be one of the first libraries to support static tensor shape type checking (with some nice opportunities to contribute to Pyrefly along the way). If you're working on 3D object detection research / training I highly recommend you give it a try!