Merge pull request #165 from Visual-Behavior/add_pose_label

Noisy aug, pose update, depth abs, render, batch_list

alodataset/transforms.py

@@ -622,13 +622,14 @@ def set_params(self):
     def apply(self, frame: Frame):
         n_frame = frame.norm01()
 
-        gaussian_noise = torch.normal(mean=0, std=self.gaussian_std, size=frame.shape)
-        shot_noise = torch.normal(mean=0, std=self.shot_std, size=frame.shape)
+        gaussian_noise = torch.normal(mean=0, std=self.gaussian_std, size=frame.shape, device=frame.device)
+        shot_noise = torch.normal(mean=0, std=self.shot_std, size=frame.shape, device=frame.device)
         noisy_frame = n_frame + n_frame * n_frame * shot_noise + gaussian_noise
         noisy_frame = torch.clip(noisy_frame, 0, 1)
 
-        if n_frame.normalization != frame.normalization:
-            n_frame = n_frame.norm_as(frame)
+        if noisy_frame.normalization != frame.normalization:
+            noisy_frame = noisy_frame.norm_as(frame)
+
         return noisy_frame
 
 

aloscene/__init__.py

@@ -8,7 +8,62 @@
 from .points_2d import Points2D
 from .points_3d import Points3D
 from .disparity import Disparity
+from .pose import Pose
 from .bounding_boxes_2d import BoundingBoxes2D
 from .bounding_boxes_3d import BoundingBoxes3D
 from .oriented_boxes_2d import OrientedBoxes2D
 from .frame import Frame
+from .tensors.spatial_augmented_tensor import SpatialAugmentedTensor
+
+from .renderer import Renderer
+
+def batch_list(tensors):
+    return SpatialAugmentedTensor.batch_list(tensors)
+
+_renderer = None
+def render(
+        views: list,
+        renderer: str = "cv",
+        size=None,
+        record_file: str = None,
+        fps=30,
+        grid_size=None,
+        skip_views=False,
+    ):
+    """Render a list of view.
+
+    Parameters
+    ----------
+    views : list
+        List of np.darray to display
+    renderer : str
+        String to set the renderer to use. Can be either ("cv" or "matplotlib")
+    cell_grid_size : tuple
+        Tuple or None. If not None, the tuple values (height, width) will be used
+        to set the size of the each grid cell of the display. If only one view is used,
+        the view will be resize to the cell grid size.
+    record_file : str
+        None by default. Used to save the rendering into one video.
+    skip_views : bool, optional
+        Skip views, in order to speed up the render process, by default False
+    """
+    global _renderer
+    _renderer = Renderer() if _renderer is None else _renderer
+
+    _renderer.render(
+        views=views,
+        renderer=renderer,
+        cell_grid_size=size,
+        record_file=record_file,
+        fps=fps,
+        grid_size=grid_size,
+        skip_views=skip_views
+    )
+
+
+def save_renderer():
+    """If render() was called with a `record_file`, then this method will
+    save the final video on the system. Warning: It is currently not possible
+    to save multiple stream directly with aloception. Todo so, one can manually create multiple `Renderer`.
+    """
+    _renderer.save()

aloscene/camera_calib.py

@@ -187,6 +187,23 @@ def __new__(cls, x, *args, **kwargs):
         tensor = super().__new__(cls, x, *args, **kwargs)
         return tensor
 
+    def translation_with(self, tgt_pos):
+        """ Compute the translation with an other pos
+
+        Parameters
+        ----------
+        tgt_pos: aloscene.Pose
+            Target pose to compute the translation with
+
+        Returns
+        -------
+        n_pos: torch.tensor
+            Translation tensor of shape (..., 3)
+        """
+        Ttgt2self = torch.linalg.solve(self.as_tensor(), tgt_pos.as_tensor())
+        translation = Ttgt2self[..., :3, -1]
+        return translation
+
     def __init__(self, x, *args, **kwargs):
         assert x.shape[-2] == 4 and x.shape[-1] == 4
         super().__init__(x)

aloscene/depth.py

@@ -23,7 +23,7 @@ class Depth(aloscene.tensors.SpatialAugmentedTensor):
     occlusion : aloscene.Mask
         Occlusion mask for this Depth map. Default value : None.
     is_bsolute: bool
-        Either depth values refer to real values or shifted and scaled ones. 
+        Either depth values refer to real values or shifted and scaled ones.
     scale: float
         Scale used to to shift depth. Pass this argument only if is_bsolute is set to True
     shift: float
@@ -32,14 +32,14 @@ class Depth(aloscene.tensors.SpatialAugmentedTensor):
 
     @staticmethod
     def __new__(
-            cls, 
-            x, 
-            occlusion: Mask = None, 
-            is_absolute=False, 
-            scale=None, 
-            shift=None, 
-            *args, 
-            names=("C", "H", "W"), 
+            cls,
+            x,
+            occlusion: Mask = None,
+            is_absolute=False,
+            scale=None,
+            shift=None,
+            *args,
+            names=("C", "H", "W"),
             **kwargs):
         if is_absolute and not (shift and scale):
             raise AttributeError('absolute depth requires shift and scale arguments')
@@ -58,9 +58,20 @@ def __new__(
     def __init__(self, x, *args, **kwargs):
         super().__init__(x)
 
-    def encode_inverse(self):
+    def encode_inverse(self, prior_clamp_min=None, prior_clamp_max=None, post_clamp_min=None, post_clamp_max=None):
         """Undo encode_absolute tansformation
-        
+
+        Parameters
+        ----------
+        prior_clamp_min: float | None
+            Clamp min depth before to convert to idepth
+        prior_clamp_max: float | None
+            Clamp max depth before to convert to idepth
+        post_clamp_min: float | None
+            Clamp min output idepth
+        post_clamp_max: float | None
+            Clamp max output idepth
+
         Exemples
         -------
         >>> not_absolute_depth = Depth(torch.ones((1, 1, 1)), is_absolute=False)
@@ -72,23 +83,40 @@ def encode_inverse(self):
         depth = self
         if not depth.is_absolute:
             raise ExecError('can not inverse depth, already inversed')
+        shift = depth.shift if depth.shift is not None else 0
+        scale = depth.scale if depth.scale is not None else 1
+
+        if prior_clamp_min is not None or prior_clamp_max is not None:
+            depth = torch.clamp(depth, min=prior_clamp_min, max=prior_clamp_max)
+
         depth = 1 / depth
-        depth = (depth - depth.shift) / depth.scale
+        depth = (depth - shift) / scale
+
+        if post_clamp_min is not None or post_clamp_max is not None:
+            depth = torch.clamp(depth, min=post_clamp_min, max=post_clamp_max)
+
         depth.scale = None
         depth.shift = None
         depth.is_absolute = False
         return depth
-    
-    def encode_absolute(self, scale=1, shift=0):
+
+    def encode_absolute(self, scale=1, shift=0, prior_clamp_min=None, prior_clamp_max=None, post_clamp_min=None, post_clamp_max=None):
         """Transforms inverted depth to absolute depth
-        
+
         Parameters
         ----------
             scale: (: float)
                 Multiplication factor. Default is 1.
-            
             shift: (: float)
                 Addition intercept. Default is 0.
+            prior_clamp_min: float | None
+                Clamp min idepth before to convert to depth
+            prior_clamp_max: float | None
+                Clamp max idepth before to convert to depth
+            post_clamp_min: float | None
+                Clamp min output idepth
+            post_clamp_max: float | None
+                Clamp max output idepth
 
         Exemples
         --------
@@ -100,12 +128,24 @@ def encode_absolute(self, scale=1, shift=0):
         depth, names = self.rename(None), self.names
         if depth.is_absolute:
             raise ExecError('depth already in absolute state, call encode_inverse first')
+
         depth = depth * scale + shift
+
+        if prior_clamp_min is not None or prior_clamp_max is not None:
+            depth = torch.clamp(depth, min=prior_clamp_min, max=prior_clamp_max)
+
         depth[torch.unsqueeze(depth < 1e-8, dim=0)] = 1e-8
         depth.scale = scale
         depth.shift = shift
         depth.is_absolute = True
-        return (1 / depth).rename(*names)
+
+        n_depth = (1 / depth).rename(*names)
+
+        if post_clamp_min is not None or post_clamp_max is not None:
+            n_depth = torch.clamp(n_depth, min=post_clamp_min, max=post_clamp_max)
+
+        return n_depth
+
 
     def append_occlusion(self, occlusion: Mask, name: str = None):
         """Attach an occlusion mask to the depth tensor.

aloscene/frame.py

@@ -6,7 +6,7 @@
 
 import aloscene
 from aloscene.renderer import View
-from aloscene import BoundingBoxes2D, BoundingBoxes3D, Depth, Disparity, Flow, Mask, Labels, Points2D, Points3D
+from aloscene import BoundingBoxes2D, BoundingBoxes3D, Depth, Disparity, Flow, Mask, Labels, Points2D, Points3D, Pose
 
 # from aloscene.camera_calib import CameraExtrinsic, CameraIntrinsic
 from aloscene.io.image import load_image
@@ -113,6 +113,7 @@ def __new__(
         tensor.add_child("depth", depth, align_dim=["B", "T"], mergeable=True)
         tensor.add_child("segmentation", segmentation, align_dim=["B", "T"], mergeable=False)
         tensor.add_child("labels", labels, align_dim=["B", "T"], mergeable=True)
+        tensor.add_child("pose", labels, align_dim=["B", "T"], mergeable=True)
 
         # Add other tensor property
         tensor.add_property("normalization", normalization)
@@ -304,6 +305,19 @@ def append_segmentation(self, segmentation: Mask, name: str = None):
         """
         self._append_child("segmentation", segmentation, name)
 
+    def append_pose(self, pose: Pose, name: str = None):
+        """Attach a pose to the frame.
+
+        Parameters
+        ----------
+        pose : :mod:`Pose <aloscene.pose>`
+            Depth to attach to the Frame
+        name : str
+            If none, the pose will be attached without name (if possible). Otherwise if no other unnamed
+            pose is attached to the frame, the pose will be added to the set of poses.
+        """
+        self._append_child("pose", pose, name)
+
     @staticmethod
     def _get_mean_std_tensor(shape, names, mean_std: tuple, device="cpu"):
         """Utils method to a get the mean and the std

aloscene/pose.py

@@ -0,0 +1,38 @@
+from aloscene.camera_calib import CameraExtrinsic
+import torch
+
+
+class Pose(CameraExtrinsic):
+    """Pose Tensor. Usually use to store World2Frame coordinates
+
+    Parameters
+    ----------
+    x: torch.Tensor
+        Pose matrix
+    """
+
+    @staticmethod
+    def __new__(cls, x, *args, names=(None, None), **kwargs):
+        tensor = super().__new__(cls, x, *args, names=names, **kwargs)
+        return tensor
+
+    def __init__(self, x, *args, **kwargs):
+        super().__init__(x)
+
+    def _hflip(self, *args, **kwargs):
+        return self.clone()
+
+    def _vflip(self, *args, **kwargs):
+        return self.clone()
+
+    def _resize(self, *args, **kwargs):
+        # Resize image does not change cam extrinsic
+        return self.clone()
+
+    def _crop(self, *args, **kwargs):
+        # Cropping image does not change cam extrinsic
+        return self.clone()
+
+    def _pad(self, *args, **kwargs):
+        # Padding image does not change cam extrinsic
+        return self.clone()