fix(rerun): stabilize Go2 3D detection previews

dimos/perception/detection/module3D.py

@@ -30,9 +30,12 @@
 from dimos.perception.detection.type.detection2d.imageDetections2D import ImageDetections2D
 from dimos.perception.detection.type.detection3d.imageDetections3DPC import ImageDetections3DPC
 from dimos.perception.detection.type.detection3d.pointcloud import Detection3DPC
+from dimos.perception.detection.visualization import letterbox_image
 from dimos.types.timestamped import align_timestamped
 from dimos.utils.reactive import backpressure
 
+_VISUALIZATION_CROP_SIZE = (320, 320)
+
 
 class Detection3DModule(Detection2DModule):
     color_image: In[Image]
@@ -45,6 +48,12 @@ class Detection3DModule(Detection2DModule):
     detected_pointcloud_1: Out[PointCloud2]
     detected_pointcloud_2: Out[PointCloud2]
 
+    # Crops paired with the successful 3D results above. These are distinct from
+    # detected_image_*, which are ranked before failed 3D projections are removed.
+    detected_3d_image_0: Out[Image]
+    detected_3d_image_1: Out[Image]
+    detected_3d_image_2: Out[Image]
+
     # just for visualization, emits latest top 3 detections in a frame
     detected_image_0: Out[Image]
     detected_image_1: Out[Image]
@@ -183,6 +192,14 @@ def _publish_detections(self, detections: ImageDetections3DPC) -> None:
         if not detections:
             return
 
+        width, height = _VISUALIZATION_CROP_SIZE
         for index, detection in enumerate(detections[:3]):
+            # Publish the crop first so selecting the following pointcloud event
+            # sees both components on the shared Rerun entity.
+            image_topic = getattr(self, "detected_3d_image_" + str(index))
+            crop = letterbox_image(detection.cropped_image(), width, height)
+            crop.frame_id = ""  # UI-only: do not attach the crop to the 3D TF tree.
+            image_topic.publish(crop)
+
             pointcloud_topic = getattr(self, "detected_pointcloud_" + str(index))
             pointcloud_topic.publish(detection.pointcloud)

dimos/perception/detection/visualization.py

@@ -0,0 +1,46 @@
+# Copyright 2025-2026 Dimensional Inc.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import numpy as np
+
+from dimos.msgs.sensor_msgs.Image import Image
+
+
+def letterbox_image(image: Image, width: int, height: int) -> Image:
+    """Create a fixed-size preview while preserving the image's aspect ratio.
+
+    DimOS sends color images to Rerun as JPEG ``EncodedImage`` values. Detection
+    crops naturally change dimensions between frames, but reusing one Rerun
+    entity for those varying dimensions triggers ``Detected change of video
+    encoding properties over time`` and can panic the Viewer. Letterboxing keeps
+    the encoded dimensions stable without stretching the crop.
+    """
+    if width <= 0 or height <= 0:
+        raise ValueError("letterbox dimensions must be positive")
+
+    scale = min(width / image.width, height / image.height)
+    resized_width = max(1, min(width, round(image.width * scale)))
+    resized_height = max(1, min(height, round(image.height * scale)))
+    resized = image.resize(resized_width, resized_height)
+
+    output_shape = (height, width, *resized.data.shape[2:])
+    output = np.zeros(output_shape, dtype=resized.data.dtype)
+    x_offset = (width - resized_width) // 2
+    y_offset = (height - resized_height) // 2
+    output[
+        y_offset : y_offset + resized_height,
+        x_offset : x_offset + resized_width,
+        ...,
+    ] = resized.data
+    return Image(data=output, format=image.format, frame_id=image.frame_id, ts=image.ts)

dimos/robot/unitree/go2/blueprints/smart/unitree_go2_detection.py

@@ -14,19 +14,56 @@
 # limitations under the License.
 
 from dimos.core.coordination.blueprints import autoconnect
+from dimos.core.global_config import global_config
 from dimos.core.transport import LCMTransport
 from dimos.msgs.sensor_msgs.Image import Image
 from dimos.msgs.sensor_msgs.PointCloud2 import PointCloud2
 from dimos.msgs.vision_msgs.Detection2DArray import Detection2DArray
 from dimos.perception.detection.module3D import Detection3DModule
+from dimos.robot.unitree.go2.blueprints.basic.unitree_go2_basic import rerun_config
 from dimos.robot.unitree.go2.blueprints.smart.unitree_go2 import unitree_go2
 from dimos.robot.unitree.go2.connection import GO2Connection
+from dimos.visualization.vis_module import vis_module
+
+
+def _topic_path(topic: object) -> str:
+    topic_str = getattr(topic, "name", None) or str(topic)
+    raw = getattr(topic, "topic", topic_str)
+    if isinstance(raw, str):
+        topic_str = raw
+    topic_str = topic_str.split("#")[0]
+    if topic_str.startswith("dimos/"):
+        topic_str = "/" + topic_str.removeprefix("dimos/")
+    elif not topic_str.startswith("/"):
+        topic_str = "/" + topic_str
+    return topic_str
+
+
+def _detection_topic_to_entity(topic: object) -> str:
+    path = _topic_path(topic)
+    parts = path.strip("/").split("/")
+    if len(parts) == 4 and parts[:2] == ["detector3d", "3d"]:
+        return f"world/detections/3d/{parts[2]}"
+    return f"world{path}"
+
+
+detection_rerun_config = {
+    **rerun_config,
+    "topic_to_entity": _detection_topic_to_entity,
+}
 
 unitree_go2_detection = (
     autoconnect(
         unitree_go2,
+        # Replace the inherited viewer bundle so its Rerun bridge uses the
+        # detection-specific topic mapping while preserving viewer="none".
+        vis_module(
+            viewer_backend=global_config.viewer,
+            rerun_config=detection_rerun_config,
+        ),
         Detection3DModule.blueprint(
             camera_info=GO2Connection.camera_info_static,
+            publish_detection_images=False,
         ),
     )
     .remappings(
@@ -37,21 +74,21 @@
     .transports(
         {
             # Detection 3D module outputs
-            ("detections", Detection3DModule): LCMTransport(
+            ("detections", Detection2DArray): LCMTransport(
                 "/detector3d/detections", Detection2DArray
             ),
-            ("detected_pointcloud_0", Detection3DModule): LCMTransport(
-                "/detector3d/pointcloud/0", PointCloud2
+            ("detected_pointcloud_0", PointCloud2): LCMTransport(
+                "/detector3d/3d/slot_0/pointcloud", PointCloud2
             ),
-            ("detected_pointcloud_1", Detection3DModule): LCMTransport(
-                "/detector3d/pointcloud/1", PointCloud2
+            ("detected_pointcloud_1", PointCloud2): LCMTransport(
+                "/detector3d/3d/slot_1/pointcloud", PointCloud2
             ),
-            ("detected_pointcloud_2", Detection3DModule): LCMTransport(
-                "/detector3d/pointcloud/2", PointCloud2
+            ("detected_pointcloud_2", PointCloud2): LCMTransport(
+                "/detector3d/3d/slot_2/pointcloud", PointCloud2
             ),
-            ("detected_image_0", Detection3DModule): LCMTransport("/detector3d/image/0", Image),
-            ("detected_image_1", Detection3DModule): LCMTransport("/detector3d/image/1", Image),
-            ("detected_image_2", Detection3DModule): LCMTransport("/detector3d/image/2", Image),
+            ("detected_3d_image_0", Image): LCMTransport("/detector3d/3d/slot_0/image", Image),
+            ("detected_3d_image_1", Image): LCMTransport("/detector3d/3d/slot_1/image", Image),
+            ("detected_3d_image_2", Image): LCMTransport("/detector3d/3d/slot_2/image", Image),
         }
     )
 )