Bug fixes, lib rename, hubconf prepared

README.md

@@ -3,7 +3,7 @@
 _[Matteo Poggi](https://mattpoggi.github.io/)\*, [Andrea Conti](https://andreaconti.github.io/)\*, [Stefano Mattoccia](http://vision.deis.unibo.it/~smatt/Site/Home.html)  *joint first authorship_
 
 
-[[arxiv]]()
+[[arxiv]](https://arxiv.org/pdf/2210.11467v1.pdf)
 [[project page]](https://andreaconti.github.io/projects/multiview_guided_multiview_stereo/)
 
 This is the official source code of Multi-View Guided Multi-View Stereo presented at [IEEE/RSJ International Conference on Intelligent Robots and Systems](https://iros2022.org/)
@@ -20,6 +20,50 @@ This is the official source code of Multi-View Guided Multi-View Stereo presente
 }
 ```
 
+## Load pretrained models and evaluate
+
+We release many of the mvs networks tested in the paper trained on Blended-MVG or on Blended-MVG and fine-tuned on DTU, with and without sparse depth points. To load these models can be simply used the `torch.hub` API.
+
+```python
+model = torch.hub.load(
+    "andreaconti/multi-view-guided-multi-view-stereo",
+    "mvsnet",               # mvsnet | ucsnet | d2hc_rmvsnet | patchmatchnet | cas_mvsnet
+    pretrained=True,
+    dataset="blended_mvg",  # blended_mvg | dtu_yao_blended_mvg
+    hints="not_guided",     # mvguided_filtered | not_guided | guided | mvguided
+)
+```
+
+Once loaded each model have the same following interface, moreover each pretrained models provides its training parameters under the attribute `train_params`.
+
+```python
+depth = model(
+    images,        # B x N x 3 x H x W
+    intrinsics,    # B x N x 3 x 3
+    extrinsics,    # B x N x 4 x 4
+    depth_values,  # B x D (128 usually)
+    hints,         # B x 1 x H x W (optional)
+)
+```
+
+Finally, we provide also an interface over the datasets used as follows. In this case is required Pytorch Lightning as dependency and the dataset must be stored locally.
+
+```python
+dm = torch.hub.load(
+    "andreaconti/multi-view-guided-multi-view-stereo",
+    "blended-mvg",  # blended_mvg | blended_mvs | dtu
+    root="data/blended-mvg",
+    hints="not_guided",     # mvguided_filtered | not_guided | guided | mvguided
+    hints_density=0.03,
+dm.prepare_data()
+dm.setup()
+dl = dm.train_dataloader()
+)
+
+```
+
+In [results.ipynb](/~https://github.com/andreaconti/multi-view-guided-multi-view/blob/main/results.ipynb) there is an example of how to reproduce some of the results showed in the paper through the `torch.hub` API.
+
 ## Installation
 
 Install the dependencies using Conda or [Mamba](/~https://github.com/mamba-org/mamba):

eval.py

@@ -23,11 +23,11 @@
 from torch.utils.data import DataLoader
 from tqdm import tqdm
 
-import lib.models as models
-from lib.datasets import find_dataset_def, find_scans
-from lib.datasets.dtu_utils import read_pfm, save_pfm
-from lib.datasets.sample_preprocess import MVSSampleTransform
-from lib.utils import *
+import guided_mvs_lib.models as models
+from guided_mvs_lib.datasets import find_dataset_def, find_scans
+from guided_mvs_lib.datasets.dtu_utils import read_pfm, save_pfm
+from guided_mvs_lib.datasets.sample_preprocess import MVSSampleTransform
+from guided_mvs_lib.utils import *
 
 
 def main():

guided_mvs_lib/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.1.0"

lib/datasets/__init__.py → guided_mvs_lib/datasets/__init__.py

@@ -18,6 +18,7 @@ def __init__(
         name: Literal["dtu_yao", "blended_mvs", "blended_mvg"],
         # args for the dataloader
         batch_size: int = 1,
+        num_workers: int = multiprocessing.cpu_count() // 2,
         # args for the dataset
         **kwargs,
     ):
@@ -27,6 +28,7 @@ def __init__(
 
         # dataloader args
         self.batch_size = batch_size
+        self.num_workers = num_workers
 
     def setup(self, stage: Optional[str] = None):
         if stage in ("fit", None):
@@ -40,23 +42,23 @@ def train_dataloader(self):
             self.mvs_train,
             batch_size=self.batch_size,
             shuffle=True,
-            num_workers=multiprocessing.cpu_count() // 2,
+            num_workers=self.num_workers,
         )
 
     def val_dataloader(self):
         return DataLoader(
             self.mvs_val,
             batch_size=1,
             shuffle=False,
-            num_workers=multiprocessing.cpu_count() // 2,
+            num_workers=self.num_workers,
         )
 
     def test_dataloader(self):
         return DataLoader(
             self.mvs_test,
             batch_size=1,
             shuffle=False,
-            num_workers=multiprocessing.cpu_count() // 2,
+            num_workers=self.num_workers,
         )
 
 

lib/datasets/dtu_blended_mvs.py → guided_mvs_lib/datasets/dtu_blended_mvs.py

@@ -11,9 +11,9 @@
 from plyfile import PlyData
 from torch.utils.data import Dataset
 
-import lib.datasets.blended_mvg_utils as mvg_utils
-import lib.datasets.blended_mvs_utils as mvs_utils
-import lib.datasets.dtu_utils as dtu_utils
+import guided_mvs_lib.datasets.blended_mvg_utils as mvg_utils
+import guided_mvs_lib.datasets.blended_mvs_utils as mvs_utils
+import guided_mvs_lib.datasets.dtu_utils as dtu_utils
 
 
 class MVSSample(TypedDict):
@@ -39,7 +39,7 @@ def _identity_fn(x: MVSSample) -> Dict:
 class MVSDataset(Dataset):
     def __init__(
         self,
-        name: Literal["dtu_yao", "blended_mvs", "blended_mvg", "eth3d"],
+        name: Literal["dtu_yao", "blended_mvs", "blended_mvg"],
         datapath: str,
         mode: Literal["train", "val", "test"],
         nviews: int = 5,

lib/datasets/sample_preprocess.py → guided_mvs_lib/datasets/sample_preprocess.py

@@ -7,7 +7,7 @@
 import torchvision.transforms.functional as F
 from PIL import Image
 
-from lib.datasets.dtu_blended_mvs import MVSSample
+from guided_mvs_lib.datasets.dtu_blended_mvs import MVSSample
 
 __all__ = ["MVSSampleTransform"]