Ptlflow

Added

- FastFlowNet model

Fixed

- Check if max_steps has negative value (the new standard value in newer PyTorch Lightning versions)

Added

- GMA model
- SCV model

Added

- DICL model
- LCV with RAFT model

Fixed

- Check if an input is an array before showing as an image during validation

Added

- MaskFlownet model
- Script for generating predictions for the test split of some datasets (currently Sintel and KITTI)

Fixed

- Inference uses no_grad annotation
- Threshold values for computing outlier metric and update table of results
- Some utility functions did not pass through all the keys from the input dicts

Added

- STaRFlow model
- Interactive plots with Plotly

Fixed

- Flow values when input images were resized during validation

Added

- LiteFlowNet 1, 2, and 3 models
- Validation can be run on a group of models at once

Fixed

- Add missing metadata to FlyingThings3DSubset dataset
- In some cases, the validations did not create the required output dirs
- Confidence log image is now saved correctly
- Add explicit list casting to validate.py to avoid errors in Python3.9

Breaking changes

- The `external` prefix was removed from all models. Therefore, models and checkpoint names have been changed in this release. The old checkpoints were replaced by new ones, so older versions will not be able to load pretrained weights anymore.
- F1-score is used for checking the accuracy of confidence, rather than L1-norm.