Fix origin ambiguity with carla dataset

CUSTOM/convert_any_avstack_labels.py

@@ -206,6 +206,7 @@ def process_frame(
     obj_filepath,
     calib_filepath,
     lidar_filepath,
+    nominal_height=1.8,
 ):
     # -- load information
     objs = read_objects_from_file(obj_filepath)
@@ -222,14 +223,7 @@ def process_frame(
                 filepath = "/" + filepath
 
         # get the new reference frame
-        ref_integ = calib.reference.integrate(start_at=GlobalOrigin3D)
-        x_new = np.array([*ref_integ.x[:2], 0])
-        q_new = tforms.transform_orientation(
-            [0, 0, tforms.transform_orientation(ref_integ.q, "quat", "euler")[2]],
-            "euler",
-            "quat",
-        )
-        ref_new = ReferenceFrame(x=x_new, q=q_new, reference=GlobalOrigin3D)
+        ref_new = calib.reference.get_ground_projected_reference()
 
         # perform compensation and save new file
         if not os.path.exists(filepath):
@@ -280,14 +274,27 @@ def process_frame(
             *reversed(bbox_3d.hwl),
             bbox_3d.yaw,
         ]
-        # ann_info['bbox_3d'][2] = 0  # HACK: consider the bottom of the box.
         ann_info[
             "num_lidar_pts"
         ] = 400  # HACK: make this wayyy faster by ignoring sum(maskfilters.filter_points_in_box(pc, bbox_3d.corners))
         ann_info["num_radar_pts"] = 0
         ann_info["velocity"] = obj.velocity.x[:2]
         ann_info["token"] = -1
 
+        # -- modifications:
+        # (1) box is bottom centered
+        if bbox_3d.where_is_t == "center":
+            ann_info["bbox_3d"][2] -= bbox_3d.h/2
+
+        # (2) yaw is defined as: yaw=0 --> along x, yaw=pi/2 --> along y
+        pass  # already this way
+
+        # (3) the anchor generation assumes a nominal sensor height
+        dx = nominal_height - bbox_3d.reference.x[2]
+        ann_info["bbox_3d"][2] -= dx
+
+        print(ann_info["bbox_3d"][2])
+
         # -- merge infos
         instances.append(ann_info)
 

CUSTOM/visualize_converted_dataset.py

@@ -0,0 +1,5 @@
+
+
+
+if __name__ == "__main__":
+    pass

configs/_base_/datasets/carla-3d-infrastructure.py

@@ -12,7 +12,7 @@
 class_names = ['car', 'bicycle', 'truck', 'motorcycle']
 metainfo = dict(classes=class_names)
 dataset_type = 'CarlaDataset'
-data_root = 'data/carla/infrastructure/skip_0'
+data_root = 'data/carla/infrastructure/skip_0/'  # trailing / is necessary
 # Input modality for nuScenes dataset, this is consistent with the submission
 # format which requires the information in input_modality.
 # Input modality for carla dataset

configs/_base_/datasets/carla-3d-vehicle.py

@@ -12,7 +12,7 @@
 class_names = ['car', 'bicycle', 'truck', 'motorcycle']
 metainfo = dict(classes=class_names)
 dataset_type = 'CarlaDataset'
-data_root = 'data/carla/vehicle/skip_0'
+data_root = 'data/carla/vehicle/skip_0/'  # trailing / is nessary
 # Input modality for nuScenes dataset, this is consistent with the submission
 # format which requires the information in input_modality.
 # Input modality for carla dataset

configs/_base_/models/pointpillars_hv_fpn_carla_vehicle.py

@@ -6,7 +6,8 @@
 # If point cloud range is modified, do remember to change all related
 # keys in the config.
 pc_range = [-100, -100, -5, 100, 100, 5]
-anchor_range = [-100, -100, -1.8, 100, 100, -1.8]
+nom_h = 1.8  # using a nominal height - this number must be in dataset conversion scripts
+anchor_range = [-100, -100, -nom_h, 100, 100, -nom_h] 
 num_classes = 4
 
 model = dict(

mmdet3d/datasets/carla_dataset.py

@@ -159,19 +159,19 @@ def parse_ann_info(self, info: dict) -> dict:
                 ann_info['centers_2d'] = np.zeros((0, 2), dtype=np.float32)
                 ann_info['depths'] = np.zeros((0), dtype=np.float32)
 
-        # the nuscenes box center is [0.5, 0.5, 0.5], we change it to be
-        # the same as KITTI (0.5, 0.5, 0)
+        # the carla box center is [0.5, 0.5, 0] which is the 
+        # same as KITTI
         # TODO: Unify the coordinates
         if self.load_type in ['fov_image_based', 'mv_image_based']:
             gt_bboxes_3d = CameraInstance3DBoxes(
                 ann_info['gt_bboxes_3d'],
                 box_dim=ann_info['gt_bboxes_3d'].shape[-1],
-                origin=(0.5, 0.5, 0.5))
+                origin=(0.5, 0.5, 0.0))
         else:
             gt_bboxes_3d = LiDARInstance3DBoxes(
                 ann_info['gt_bboxes_3d'],
                 box_dim=ann_info['gt_bboxes_3d'].shape[-1],
-                origin=(0.5, 0.5, 0.5)).convert_to(self.box_mode_3d)
+                origin=(0.5, 0.5, 0.0)).convert_to(self.box_mode_3d)
 
         ann_info['gt_bboxes_3d'] = gt_bboxes_3d