Torch Points3d Versions Save

1.3.0

Added

• MS-SVConv: https://arxiv.org/abs/2103.14533 (thanks @humanpose1)
• added new data generations techniques for the self-supervised learning (PeriodicSampling, IrregularSampling EllipsoidCrop) (thanks @humanpose1)
• More ETH benchmark dataset (thanks @humanpose1)

Changed

• Minkowski 0.5 support

Bug fixes

• Fix bug in data loader https://github.com/nicolas-chaulet/torch-points3d/issues/443 thanks @JloveU
• Fix bug in base unet that created problems when loading pointnet++ model checkpoint

Added

• Support for the IRALab benchmark (https://arxiv.org/abs/2003.12841), with data from the ETH, Canadian Planetary, Kaist and TUM datasets. (thanks @simone-fontana)
• Added Kitti for semantic segmentation and registration (first outdoor dataset for semantic seg)
• Possibility to load pretrained models by adding the path in the confs for finetuning.
• Lottery transform to use randomly selected transforms for data augmentation
• Batch size campling function to ensure that batches don't get too large
• TorchSparse backend for sparse convolutions
• Possibility to build sparse convolution networks with Minkowski Engine or TorchSparse
• PVCNN model for semantic segmentation (thanks @CCInc)

Bug fix

• Dataset configurations are saved in the checkpoints so that models can be created without requiring the actual dataset
• Trainer was giving a warning for models that could not be re created when they actually could
• BatchNorm1d fix (thanks @Wundersam)
• Fix process hanging when processing scannet with multiprocessing (thanks @zetyquickly)
• wandb does not log the weights when set in private mode (thanks @jamesjiro)
• Fixed VoteNet loss definitions and data augmentation parameters (got up to 59.2% mAP25)

Changed

• More general API for Minkowski with support for Bottleneck blocks and Squeeze and excite.
• Docker images tags on dockerhub are now latest-gpu and latest-cpu for the latest CPU adn GPU images.

Removed

• Removed VoteNet from the API because it was not up to date. You can still use the models defined there

1.1.1

Added

• Teaser support for registration
• Examples for using pretrained registration models

Changed

• Moved to PyTorch 1.6 as officially supported PyTorch version

Bug fix

• Add context = ssl._create_unverified_context(), data = urllib.request.urlopen(url, context=context) within download_ulr, so ModelNet and ShapeNet can download.

This release brings a few bug fixes as well as some new feature:

• Panoptic segmentation with PointGroup
• Object detection with VoteNet
• Registry of pre trained models on s3dis, 3dmatch and kitti
• Windows support
• Docker image to run trainings

Thanks a lot to all contributors @Uakh @humanpose1 @loicland @CCInc @tchaton

Changed

• We now support the latest PyTorch
• Migration to the latest PyTorch Geometric and dependencies

Bugfixes

• #273 (support python 3.7)

v1 is here!

Bugfix

• Pre transform is being overriden by the inference transform

Added

• Customizable number of channels at the output of the API models
• API models expose output number of channels as a property
• Added Encoder to the API
• Sampled ModelNet dataset for point clouds