Plotly subsampling fix

Summary: When viewing two or more pointclouds in a single plot, we should be subsampling each one separately rather than subsampling their union.

Reviewed By: nikhilaravi

Differential Revision: D27010770

fbshipit-source-id: 3c7e04a6049edd39756047f985d5a82c2601b3a2

pytorch3d/vis/plotly_vis.py

@@ -61,14 +61,14 @@ def plot_scene(
     **kwargs,
 ):
     """
-    Main function to visualize Meshes and Pointclouds.
+    Main function to visualize Meshes, Cameras and Pointclouds.
     Plots input Pointclouds, Meshes, and Cameras data into named subplots,
     with named traces based on the dictionary keys. Cameras are
     rendered at the camera center location using a wireframe.
 
     Args:
         plots: A dict containing subplot and trace names,
-            as well as the Meshes and Pointclouds objects to be rendered.
+            as well as the Meshes, Cameras and Pointclouds objects to be rendered.
             See below for examples of the format.
         viewpoint_cameras: an instance of a Cameras object providing a location
             to view the plotly plot from. If the batch size is equal
@@ -574,11 +574,21 @@ def _add_pointcloud_trace(
     pointclouds = pointclouds.detach().cpu()
     verts = pointclouds.points_packed()
     features = pointclouds.features_packed()
-    total_points_count = max_points_per_pointcloud * len(pointclouds)
 
     indices = None
-    if verts.shape[0] > total_points_count:
-        indices = np.random.choice(verts.shape[0], total_points_count, replace=False)
+    if pointclouds.num_points_per_cloud().max() > max_points_per_pointcloud:
+        start_index = 0
+        index_list = []
+        for num_points in pointclouds.num_points_per_cloud():
+            if num_points > max_points_per_pointcloud:
+                indices_cloud = np.random.choice(
+                    num_points, max_points_per_pointcloud, replace=False
+                )
+                index_list.append(start_index + indices_cloud)
+            else:
+                index_list.append(start_index + np.arange(num_points))
+            start_index += num_points
+        indices = np.concatenate(index_list)
         verts = verts[indices]
 
     color = None