R2D2 Feature extractor #46
Conversation
hloc/extractors/r2d2.py
Outdated
| RGB_mean = torch.cuda.FloatTensor([0.485, 0.456, 0.406]) | ||
| RGB_std = torch.cuda.FloatTensor([0.229, 0.224, 0.225]) | ||
| norm_RGB = tvf.Normalize(mean=RGB_mean, std=RGB_std) |
There was a problem hiding this comment.
- This creates tensors on GPU even if this module is merely imported.
- hloc should not break on GPU-only machines.
I suggest to:
- create
meanandstdin_initastorch.tensor(CPU) and register them withregister_buffer. This way, they will be automatically moved to GPU with the model - call
tvf.functional.normalizein_forward(meanandstdwill already be on the correct device)
There was a problem hiding this comment.
I pass the mean and std as list as required in this function(https://pytorch.org/docs/stable/_modules/torchvision/transforms/functional.html#normalize) and let the function itself to move the tensors to the suitable device. Is that ok?
hloc/extractors/r2d2.py
Outdated
| 'top-k': 5000, | ||
|
|
||
| 'scale-f': 2**0.25, | ||
| 'min-size': 256, | ||
| 'max-size': 1024, | ||
| 'min-scale': 0, | ||
| 'max-scale': 1, | ||
|
|
||
| 'reliability-thr': 0.7, | ||
| 'repetability-thr': 0.7, |
There was a problem hiding this comment.
- I usually use underscored instead of dashes for config entries
- Could we replace
top-kwithmax_keypointsto be consistent withsuperpoint.py scale-fis not very explicit, is itscale_factor?thr-->threshold
There was a problem hiding this comment.
These variable names are set to be consistent with the original r2d2 repo. But I believe your suggestion of making it consistent with hloc is more reasonable.
hloc/extractors/r2d2.py
Outdated
|
|
||
| def _forward(self, data): | ||
| img = data['image'] | ||
| img = norm_RGB(img[0])[None] |
There was a problem hiding this comment.
tvf normalization supports batching
hloc/extractors/r2d2.py
Outdated
| self.detector = NonMaxSuppression(rel_thr=conf['reliability-thr'], | ||
| rep_thr=conf['repetability-thr']) |
There was a problem hiding this comment.
This does not seem to respect PEP8
…tore the tvf.Normalize function norm_rgb as a member function in the class. The device will be take care in tvf.functional.Normalize
… keypoints/descriptors are the same). Successfully run on CPU but the results seem not to match very well
|
@lxxue How did you handle the descriptor dimension when using superglue for matching? |
|
@Matstah Sorry it doesn't support matching with superglue as they are trained on different features of different dimensions. SuperGlue is descriptor-specific so you indeed need to retrain for each new descriptor. |
|
Thanks a lot for the changes! |
No description provided.