Merge pull request #962 from StanfordVL/fix/key-array-caching

Fix Key Array Caching

omnigibson/sensors/vision_sensor.py

@@ -329,8 +329,6 @@ def _remap_modality(self, modality, obs, info, raw_obs):
             obs[modality], info[modality] = self._remap_instance_segmentation(
                 obs[modality],
                 id_to_labels,
-                obs["seg_semantic"],
-                info["seg_semantic"],
                 id=(modality == "seg_instance_id"),
             )
         elif "bbox" in modality:
@@ -387,16 +385,14 @@ def _remap_semantic_segmentation(self, img, id_to_labels):
 
         return VisionSensor.SEMANTIC_REMAPPER.remap(replicator_mapping, semantic_class_id_to_name(), img, image_keys)
 
-    def _remap_instance_segmentation(self, img, id_to_labels, semantic_img, semantic_labels, id=False):
+    def _remap_instance_segmentation(self, img, id_to_labels, id=False):
         """
         Remap the instance segmentation image to our own instance IDs.
         Also, correct the id_to_labels input with our new labels and return it.
 
         Args:
             img (th.tensor): Instance segmentation image to remap
             id_to_labels (dict): Dictionary of instance IDs to class labels
-            semantic_img (th.tensor): Semantic segmentation image to use for instance registry
-            semantic_labels (dict): Dictionary of semantic IDs to class labels
             id (bool): Whether to remap for instance ID segmentation
         Returns:
             th.tensor: Remapped instance segmentation image

omnigibson/utils/vision_utils.py

@@ -71,12 +71,14 @@ class Remapper:
     def __init__(self):
         self.key_array = th.empty(0, dtype=th.int32, device="cuda")  # Initialize the key_array as empty
         self.known_ids = set()
+        self.unlabelled_ids = set()
         self.warning_printed = set()
 
     def clear(self):
         """Resets the key_array to empty."""
         self.key_array = th.empty(0, dtype=th.int32, device="cuda")
         self.known_ids = set()
+        self.unlabelled_ids = set()
 
     def remap(self, old_mapping, new_mapping, image, image_keys=None):
         """
@@ -109,6 +111,15 @@ def remap(self, old_mapping, new_mapping, image, image_keys=None):
             # Copy the previous key array into the new key array
             self.key_array[: len(prev_key_array)] = prev_key_array
 
+        # Retrospectively inspect our cached ids against the old mapping and update the key array
+        updated_ids = set()
+        for unlabelled_id in self.unlabelled_ids:
+            if unlabelled_id in old_mapping and old_mapping[unlabelled_id] != "unlabelled":
+                # If an object was previously unlabelled but now has a label, we need to update the key array
+                updated_ids.add(unlabelled_id)
+        self.unlabelled_ids -= updated_ids
+        self.known_ids -= updated_ids
+
         new_keys = old_mapping.keys() - self.known_ids
         if new_keys:
             self.known_ids.update(new_keys)
@@ -118,6 +129,9 @@ def remap(self, old_mapping, new_mapping, image, image_keys=None):
                 new_key = next((k for k, v in new_mapping.items() if v == label), None)
                 assert new_key is not None, f"Could not find a new key for label {label} in new_mapping!"
                 self.key_array[key] = new_key
+                if label == "unlabelled":
+                    # Some objects in the image might be unlabelled first but later get a valid label later, so we keep track of them
+                    self.unlabelled_ids.add(key)
 
         # For all the values that exist in the image but not in old_mapping.keys(), we map them to whichever key in
         # new_mapping that equals to 'unlabelled'. This is needed because some values in the image don't necessarily