feat(BEVFusion-Camera): release BEVFusion-camera 2.7.x

autoware_ml/configs/detection3d/dataset/t4dataset/base.py

@@ -29,6 +29,9 @@
     "db_j6gen2_v7",
     "db_j6gen2_v8",
     "db_j6gen2_v9",
+    "db_j6gen2_v10",
+    "db_j6gen2_v11",
+    "db_j6gen2_v12",
     "db_largebus_v1",
     "db_largebus_v2",
     "db_largebus_v3",
@@ -146,6 +149,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 class_names = ["car", "truck", "bus", "bicycle", "pedestrian", "traffic_cone", "barrier"]

autoware_ml/configs/detection3d/dataset/t4dataset/j6gen2.py

@@ -27,6 +27,9 @@
     "db_j6gen2_v7",
     "db_j6gen2_v8",
     "db_j6gen2_v9",
+    "db_j6gen2_v10",
+    "db_j6gen2_v11",
+    "db_j6gen2_v12",
 ]
 
 dataset_test_groups = {
@@ -127,6 +130,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 class_names = [

autoware_ml/configs/detection3d/dataset/t4dataset/j6gen2_base.py

@@ -27,6 +27,9 @@
     "db_j6gen2_v7",
     "db_j6gen2_v8",
     "db_j6gen2_v9",
+    "db_j6gen2_v10",
+    "db_j6gen2_v11",
+    "db_j6gen2_v12",
     "db_largebus_v1",
     "db_largebus_v2",
     "db_largebus_v3",
@@ -133,6 +136,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 

autoware_ml/configs/detection3d/dataset/t4dataset/jpntaxi_base.py

@@ -123,6 +123,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 class_names = [

autoware_ml/configs/detection3d/dataset/t4dataset/jpntaxi_gen2.py

@@ -120,6 +120,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 class_names = [

autoware_ml/configs/detection3d/dataset/t4dataset/largebus.py

@@ -122,6 +122,8 @@
     "traffic_cone": "traffic_cone",
     "trafficcone": "traffic_cone",
     "barrier": "barrier",
+    "other_vehicle": "car",
+    "other_pedestrian": "pedestrian",
 }
 
 class_names = [

autoware_ml/configs/detection3d/default_runtime.py

@@ -2,9 +2,17 @@
 
 default_hooks = dict(
     timer=dict(type="IterTimerHook"),
-    logger=dict(type="LoggerHook", interval=50),
+    logger=dict(
+        type="LoggerHook",
+        interval=50,
+        backend_args=dict(backend="local"),
+    ),
     param_scheduler=dict(type="ParamSchedulerHook"),
-    checkpoint=dict(type="CheckpointHook", interval=-1),
+    checkpoint=dict(
+        type="CheckpointHook",
+        interval=-1,
+        backend_args=dict(backend="local"),
+    ),
     sampler_seed=dict(type="DistSamplerSeedHook"),
     visualization=dict(type="Det3DVisualizationHook"),
 )

autoware_ml/detection3d/datasets/t4dataset.py

@@ -2,6 +2,7 @@
 from typing import List
 
 import numpy as np
+import tqdm
 from mmdet3d.datasets import NuScenesDataset
 from mmengine.logging import print_log
 from mmengine.registry import DATASETS
@@ -56,16 +57,14 @@ def filter_data(self) -> List[dict]:
             return self.data_list
 
         filtered_data_list = []
-        for entry in self.data_list:
+        for entry in tqdm.tqdm(self.data_list, desc="Filtering data"):
             filtered = False
             for camera_order in filter_frames_with_camera_order:
                 if camera_order not in entry["images"]:
                     filtered = True
                     break
 
-                if entry["images"][camera_order]["img_path"] is None or not osp.exists(
-                    entry["images"][camera_order]["img_path"]
-                ):
+                if entry["images"][camera_order]["img_path"] is None:
                     filtered = True
                     break
 
@@ -180,6 +179,7 @@ def parse_data_info(self, info: dict) -> dict:
                             cam_prefix,
                             img_info["img_path"],
                         )
+                    # print_log(f"Camera path: {img_info['img_path']}", logger="current")
 
             if self.default_cam_key is not None:
                 info["img_path"] = info["images"][self.default_cam_key]["img_path"]
@@ -192,4 +192,7 @@ def parse_data_info(self, info: dict) -> dict:
                 else:
                     info["lidar2img"] = info["cam2img"] @ info["lidar2cam"]
 
+        # Default difficulty to 0 if not present
+        if "difficulty" not in info:
+            info["difficulty"] = 0
         return info

autoware_ml/detection3d/datasets/transforms/__init__.py

@@ -1,3 +1,4 @@
+from .local_3d_bbox import Local3DBBoxExpand
 from .object_min_points_filter import ObjectMinPointsFilter
 
-__all__ = ["ObjectMinPointsFilter"]
+__all__ = ["ObjectMinPointsFilter", "Local3DBBoxExpand"]

autoware_ml/detection3d/datasets/transforms/local_3d_bbox.py

@@ -0,0 +1,58 @@
+from typing import List
+
+import numpy as np
+from mmcv.transforms import BaseTransform
+from mmdet3d.structures.ops import box_np_ops
+from mmengine.registry import TRANSFORMS
+
+
+@TRANSFORMS.register_module()
+class Local3DBBoxExpand(BaseTransform):
+    """Locally expand the 3D bounding boxes by scaling the width, which it doesn't scale the points.
+
+    Args:
+        expand_widths: (List[float]): Uniformly sampled expand width.
+        width_dim: (int): The dimension of the width. Default is 4, which is the width dimension of the 3D
+                  bounding box. Since 3D Bbox is in the format of [x, y, z, dx, dy, dz, heading], the width dimension is the
+                    4th dimension.
+            label_ids: (List[int]): The label IDs to expand. If None, all label IDs will be expanded.
+    """
+
+    def __init__(self, expand_widths: List[float], width_dim: int = 4, label_ids: List[int] = None) -> None:
+        assert isinstance(expand_widths, list)
+        assert len(expand_widths) == 2
+        assert expand_widths[0] < expand_widths[1]
+        self.expand_widths = expand_widths
+        self.width_dim = width_dim
+        self.label_ids = label_ids
+
+    def transform(self, input_dict: dict) -> dict:
+        """Call function to locally augment the 3D bounding boxes by scaling the width.
+
+        Args:
+            input_dict (dict): Result dict from loading pipeline.
+
+        Returns:
+            dict: Results after locally augmenting the 3D bounding boxes by scaling the width, 'gt_bboxes_3d' \
+                key is updated in the result dict.
+        """
+        # Label mask
+        if self.label_ids is not None:
+            label_masks = [True if label in self.label_ids else False for label in input_dict["gt_labels_3d"]]
+        else:
+            label_masks = np.ones(len(input_dict["gt_labels_3d"]), dtype=bool)
+
+        for i in range(len(input_dict["gt_bboxes_3d"])):
+            if not label_masks[i]:
+                continue
+
+            expand_width = np.random.uniform(self.expand_widths[0], self.expand_widths[1])
+            input_dict["gt_bboxes_3d"].tensor[i, self.width_dim] += expand_width
+
+        return input_dict
+
+    def __repr__(self) -> str:
+        """str: Return a string that describes the module."""
+        repr_str = self.__class__.__name__
+        repr_str += f"(expand_widths={self.expand_widths}, width_dim={self.width_dim}, label_ids={self.label_ids})"
+        return repr_str

projects/BEVFusion/bevfusion/__init__.py

@@ -1,11 +1,19 @@
 from .bevfusion import BEVFusion
 from .bevfusion_head import BEVFusionHead, ConvFuser
 from .bevfusion_necks import GeneralizedLSSFPN
+from .bevfusion_voxel_encoder import BEVFusionVoxelFeatureNet, HardSimpleVoxelSinCosEncoder
 from .depth_lss import DepthLSSTransform, LSSTransform
-from .loading import BEVLoadMultiViewImageFromFiles
+from .depth_lss_v2 import LSSTransformV2, LSSTransformV2DepthAware
+from .loading import BEVLoadMultiViewImageFromFiles, PointsToMultiViewImageDepths
 from .sparse_encoder import BEVFusionSparseEncoder
 from .transformer import TransformerDecoderLayer
-from .transforms_3d import BEVFusionGlobalRotScaleTrans, BEVFusionRandomFlip3D, GridMask, ImageAug3D
+from .transforms_3d import (
+    BEVFusionGlobalRotScaleTrans,
+    BEVFusionRandomFlip3D,
+    BEVFusionRemoveLiDARPoints,
+    GridMask,
+    ImageAug3D,
+)
 from .utils import BBoxBEVL1Cost, HeuristicAssigner3D, HungarianAssigner3D, IoU3DCost, TransFusionBBoxCoder
 
 __all__ = [
@@ -26,5 +34,11 @@
     "TransformerDecoderLayer",
     "BEVFusionRandomFlip3D",
     "BEVFusionGlobalRotScaleTrans",
+    "BEVFusionRemoveLiDARPoints",
     "TransFusionBBoxCoder",
+    "HardSimpleVoxelSinCosEncoder",
+    "BEVFusionVoxelFeatureNet",
+    "LSSTransformV2",
+    "PointsToMultiViewImageDepths",
+    "LSSTransformV2DepthAware",
 ]