This is a major update and introduces breaking changes to v0.2.
The list of changes below is not comprehensive, there may be other changes that are not listed.
Added
- New models:
- DIP
- Flow1D
- FlowFormer++
- GMFlow+
- MatchFlow
- MS-RAFT+
- RPKNet
-SeparableFlow
- SKFlow
- VideoFlow
- New datasets:
- Middlebury
- Monkaa
- Spring
- Option to use RAFT's alt_cuda_corr for supported models
- Also added a pure PyTorch version of alt_cuda_corr, which is slower but does not need to be compiled
Fixed
- Compatibility with PyTorch 2.0, 2.1
- Compatibility issues with PyTorch Lightning 1.9
- Resizing augmentation when the valid mask is sparse
- Models should produce results more similar to the paper
- HOWEVER, we do not guarantee our results are correct. Use the official implementations for the most accurate results.
Changed
- Each model now handle its own IO reshaping, instead of using padding for all models
- speed_benchmark.py becomes model_benchmark.py and records more metrics
- Renamed model: pwcdcnet -> pwcnet, pwcnet -> pwcnet_nodc
- Updated requirements, support for many old versions are dropped
- Important requirements:
- python>=3.8,<3.12
- torch>=1.13,<2.2
- lightning>=1.9.0,<2