GitLab

update back to original version of torch.div and torch.meshgrid to also support lower version pytorch

* support PointPillar (anchor-based, anchor-free) for WOD datasets

* update OpenPCDet to v0.5.1

bugfixed: weights transformation between spconv 1.x and spconv 2.x, should consider both native/implicit spconv 2.x

- as per the spconv documentation, the order of channels in weights is now changed
- we need to get all the affected weights, using the util function in spconv_utils
- we need to check the version with the version checker in __init__.py
- we need to transpose last two dimensions of existing weights, from older models
- to avoid duplication a single _load_state_dict method is added

missing kornia. before image_to_tensor

* add voxel roi pooling

* add voxel r-cnn head

* add voxel query

* add voxel r-cnn

* add Voxel R-CNN

* add infos about Voxel R-CNN

* add voxel_rcnn_car.yaml

* add infos about Voxel R-CNN
Co-authored-by: Shaoshuai Shi <shaoshuaics@gmail.com>

* Download DeepLabV3 if not available yet

* Fix dtype issue

Fix RuntimeError: expected backend CPU and dtype Float but got backend CPU and dtype Long

* Added CaDDN detector and support for image, depth map, and 2D GT box
dataloading

* Moved image flip augmentation to augmentor_utils

* Updated default get item list to include points

* Moved utils functions into transform_utils

* Combined FFE + F2V into ImageVFE, renamed FFE to FFN, moved depth downsample into data_processor

* Updated README with updated CaDDN weights

* Updated comments for image vfe

* support SECOND-IoU

* rename file

* modify readme and fix model name in __init__.py

If there are less than half of model_cfg.NUM_KEYPOINTS it will lead to 
RuntimeError: unsupported operation: some elements of the input tensor and the written-to tensor refer to a single memory location. 
Please clone() the tensor before performing the operation.

caused by `features` in pointnet2_backbone.py:78, which is possible to be a non-contiguous tensor, being directly passed into SA_modules without calling .contiguous()

* bugfixed: ignore empty boxes in visualization

* add contiguous() after transpose() at pointrcnn_head.py