Synthia training

README.md

@@ -27,7 +27,7 @@ Modify the `BATCH_SIZE` accordingly.
 The first argument is the GPU id and the second argument is the path postfix
 and the last argument is the miscellaneous arguments.
 
-The official evaluation metric for ScanNet is mIoU. OA, Overal Accuracy is not the official metric and is a lot easier and should not be used soley for any official papers without mIoU for fair comparison.
+The official evaluation metric for ScanNet is mIoU. OA, Overal Accuracy is not the official metric similar to the 2D image semantic segmentation as it is a lot easier and doesn't reflect the quality of the semantic segmentation. This is due to the fact that most of the scenes consist of large structures such as walls, floors and these will dominate the statistics. Thus, overall accuracy should not be used solely to prevent reviewers from comparing the baselines fairly.
 
 ## Synthia 4D Experiment
 
@@ -36,10 +36,23 @@ The official evaluation metric for ScanNet is mIoU. OA, Overal Accuracy is not t
 2. Extract
 
 ```
+cd /path/to/extract/synthia4d
+wget http://cvgl.stanford.edu/data2/Synthia4D.tar
 tar -xf Synthia4D.tar
-tar -xzf *.tar.bz2
+tar -xvjf *.tar.bz2
 ```
 
+3. Training
+
+```
+export BATCH_SIZE=N; \
+./scripts/train_synthia4d.sh 0 \
+	"-default" \
+	"--synthia_path /path/to/extract/synthia4d"
+```
+
+The above script trains a network. You have to change the arguments accordingly. The first argument to the script is the GPU id. Second argument is the log directory postfix; change to mark your experimental setup. The final argument is a series of the miscellaneous aruments. You have to specify the synthia directory here. Also, you have to wrap all arguments with " ".
+
 
 ## Model Zoo
 

lib/dataset.py

@@ -258,7 +258,8 @@ def __getitem__(self, index):
       pointcloud = pointcloud[inds]
 
     # Prevoxel transformations
-    pointcloud = self.prevoxel_transform(pointcloud)
+    if self.prevoxel_transform is not None:
+      pointcloud = self.prevoxel_transform(pointcloud)
 
     coords, feats, labels = self.convert_mat2cfl(pointcloud)
     coords, feats, labels, transformation = self.voxelizer.voxelize(

lib/datasets/__init__.py

@@ -1,10 +1,12 @@
-from .synthia import SynthiaVoxelizationDataset, SynthiaTemporalVoxelizationDataset
+from .synthia import SynthiaCVPR15cmVoxelizationDataset, SynthiaCVPR30cmVoxelizationDataset, \
+    SynthiaAllSequencesVoxelizationDataset
 from .stanford import StanfordVoxelizationDataset, StanfordVoxelization2cmDataset
 from .scannet import ScannetVoxelizationDataset, ScannetVoxelization2cmDataset
 
 DATASETS = [
     StanfordVoxelizationDataset, StanfordVoxelization2cmDataset, ScannetVoxelizationDataset,
-    ScannetVoxelization2cmDataset, SynthiaVoxelizationDataset, SynthiaTemporalVoxelizationDataset
+    ScannetVoxelization2cmDataset, SynthiaCVPR15cmVoxelizationDataset,
+    SynthiaCVPR30cmVoxelizationDataset, SynthiaAllSequencesVoxelizationDataset
 ]
 
 

lib/datasets/scannet.py

@@ -99,7 +99,7 @@ def __init__(self,
     data_root = config.scannet_path
     if phase not in [DatasetPhase.Train, DatasetPhase.TrainVal]:
       self.CLIP_BOUND = self.TEST_CLIP_BOUND
-    data_paths = read_txt(os.path.join(data_root, self.DATA_PATH_FILE[phase]))
+    data_paths = read_txt(os.path.join('./splits/scannt', self.DATA_PATH_FILE[phase]))
     logging.info('Loading {}: {}'.format(self.__class__.__name__, self.DATA_PATH_FILE[phase]))
     super().__init__(
         data_paths,

lib/datasets/synthia.py

@@ -108,9 +108,10 @@ class SynthiaVoxelizationDataset(VoxelizationDataset):
   # Voxelization arguments
   CLIP_BOUND = ((-1800, 1800), (-1800, 1800), (-1800, 1800))
   TEST_CLIP_BOUND = ((-2500, 2500), (-2500, 2500), (-2500, 2500))
-  VOXEL_SIZE = 30  # cm
+  VOXEL_SIZE = 15  # cm
+
+  PREVOXELIZE_VOXEL_SIZE = 7.5
   # Elastic distortion, (granularity, magitude) pairs
-  PREVOXELIZE_VOXEL_SIZE = 30
   ELASTIC_DISTORT_PARAMS = ((80, 300),)
 
   # Augmentation arguments
@@ -124,10 +125,9 @@ class SynthiaVoxelizationDataset(VoxelizationDataset):
 
   # Split used in the Minkowski ConvNet, CVPR'19
   DATA_PATH_FILE = {
-      DatasetPhase.Train: 'train_raw_difficult.txt',
-      DatasetPhase.Val: 'val_raw_difficult.txt',
-      DatasetPhase.TrainVal: 'trainval_raw_difficult.txt',
-      DatasetPhase.Test: 'test_raw_difficult.txt'
+      DatasetPhase.Train: 'train_cvpr19.txt',
+      DatasetPhase.Val: 'val_cvpr19.txt',
+      DatasetPhase.Test: 'test_cvpr19.txt'
   }
 
   def __init__(self,
@@ -144,7 +144,7 @@ def __init__(self,
     if phase not in [DatasetPhase.Train, DatasetPhase.TrainVal]:
       self.CLIP_BOUND = self.TEST_CLIP_BOUND
     data_root = config.synthia_path
-    data_paths = read_txt(osp.join(data_root, self.DATA_PATH_FILE[phase]))
+    data_paths = read_txt(osp.join('./splits/synthia4d', self.DATA_PATH_FILE[phase]))
     data_paths = [d.split()[0] for d in data_paths]
     logging.info('Loading {}: {}'.format(self.__class__.__name__, self.DATA_PATH_FILE[phase]))
     super().__init__(
@@ -164,10 +164,11 @@ class SynthiaTemporalVoxelizationDataset(TemporalVoxelizationDataset):
   # Voxelization arguments
   CLIP_BOUND = ((-1800, 1800), (-1800, 1800), (-1800, 1800))
   TEST_CLIP_BOUND = ((-2500, 2500), (-2500, 2500), (-2500, 2500))
-  VOXEL_SIZE = 30  # cm
+  VOXEL_SIZE = 15  # cm
 
-  PREVOXELIZE_VOXEL_SIZE = 30
-  ELASTIC_DISTORT_PARAMS = ((80, 300),)
+  PREVOXELIZE_VOXEL_SIZE = 7.5
+  # For temporal sequences, the voxel locations has to be aligned exactly.
+  ELASTIC_DISTORT_PARAMS = None
 
   # Augmentation arguments
   ROTATION_AUGMENTATION_BOUND = ((0, 0), (-np.pi, np.pi), (0, 0))
@@ -250,6 +251,22 @@ def _transform(xyz, intrinsic, extrinsic):
     return ptc, center
 
 
+class SynthiaCVPR15cmVoxelizationDataset(SynthiaVoxelizationDataset):
+  pass
+
+
+class SynthiaCVPR30cmVoxelizationDataset(SynthiaVoxelizationDataset):
+  VOXEL_SIZE = 30
+
+
+class SynthiaAllSequencesVoxelizationDataset(SynthiaVoxelizationDataset):
+  DATA_PATH_FILE = {
+      DatasetPhase.Train: 'train_raw.txt',
+      DatasetPhase.Val: 'val_raw.txt',
+      DatasetPhase.Test: 'test_raw.txt'
+  }
+
+
 class TestSynthia(unittest.TestCase):
 
   @debug_on()