tier4 · KSeangTan · Feb 13, 2026 · Jan 14, 2026 · Jan 14, 2026 · Jan 20, 2026
@@ -173,6 +173,35 @@ python projects/BEVFusion/deploy/torch2onnx.py \
 
 ```
 
+To export a camera-only model, please use the following command:
+
+```bash
+DEPLOY_CFG_MAIN_BODY=configs/deploy/bevfusion_camera_point_bev_tensorrt_dynamic.py.py
+DEPLOY_CFG_IMAGE_BACKBONE=configs/deploy/bevfusion_camera_backbone_tensorrt_dynamic.py
+
+MODEL_CFG=...
+CHECKPOINT_PATH=...
+WORK_DIR=...
+
+python projects/BEVFusion/deploy/torch2onnx.py \
+  ${DEPLOY_CFG_MAIN_BODY} \
+  ${MODEL_CFG} \
+  ${CHECKPOINT_PATH} \
+  --device cuda:0 \
+  --work-dir ${WORK_DIR} \
+  --module camera_bev_only_network
+
+
+python projects/BEVFusion/deploy/torch2onnx.py \
+  ${DEPLOY_CFG_IMAGE_BACKBONE} \
+  ${MODEL_CFG} \
+  ${CHECKPOINT_PATH} \
+  --device cuda:0 \
+  --work-dir ${WORK_DIR} \
+  --module image_backbone
+```
+Note that this camera-only model takes lidar pointclouds as an input for a depth map, and we will release a model without lidar pointclouds in another release.
+
 This will generate two models in the `WORK_DIR` folder. `end2end.onnx` corresponds to the standard exported model ,whereas `end2end_fixed.onnx` contains a fix for the `TopK` operator (compatibility issues between `mmdeploy` and `TensorRT`).
 
 ## TODO

@@ -355,7 +355,7 @@ def extract_feat(
             )
             features.append(img_feature)
 
-        if points is not None and self.pts_middle_encoder is not None:
+        if self.pts_middle_encoder is not None:
             pts_feature = self.extract_pts_feat(
                 batch_inputs_dict.get("voxels", {}).get("voxels", None),
                 batch_inputs_dict.get("voxels", {}).get("coors", None),

@@ -0,0 +1,93 @@
+codebase_config = dict(type="mmdet3d", task="VoxelDetection", model_type="end2end")
+
+custom_imports = dict(
+    imports=[
+        "projects.BEVFusion.deploy",
+        "projects.BEVFusion.bevfusion",
+        "projects.SparseConvolution",
+    ],
+    allow_failed_imports=False,
+)
+
+depth_bins = 129
+# TODO(KokSeang): Read this parameter from a base config
+feature_dims = (48, 96)
+# image_dims = (640, 576)
+
+backend_config = dict(
+    type="tensorrt",
+    common_config=dict(max_workspace_size=1 << 32),
+    model_inputs=[
+        dict(
+            input_shapes=dict(
+                # TODO(TIERIV): Optimize. Now, using points will increase latency significantly
+                points=dict(min_shape=[5000, 4], opt_shape=[50000, 4], max_shape=[200000, 4]),
+                lidar2image=dict(min_shape=[1, 4, 4], opt_shape=[6, 4, 4], max_shape=[6, 4, 4]),
+                img_aug_matrix=dict(min_shape=[1, 4, 4], opt_shape=[6, 4, 4], max_shape=[6, 4, 4]),
+                geom_feats=dict(
+                    min_shape=[0 * depth_bins * feature_dims[0] * feature_dims[1], 4],
+                    opt_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1] // 2, 4],
+                    max_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1], 4],
+                ),
+                kept=dict(
+                    min_shape=[0 * depth_bins * feature_dims[0] * feature_dims[1]],
+                    opt_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1]],
+                    max_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1]],
+                ),
+                ranks=dict(
+                    min_shape=[0 * depth_bins * feature_dims[0] * feature_dims[1]],
+                    opt_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1] // 2],
+                    max_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1]],
+                ),
+                indices=dict(
+                    min_shape=[0 * depth_bins * feature_dims[0] * feature_dims[1]],
+                    opt_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1] // 2],
+                    max_shape=[6 * depth_bins * feature_dims[0] * feature_dims[1]],
+                ),
+                image_feats=dict(
+                    min_shape=[0, 256, feature_dims[0], feature_dims[1]],
+                    opt_shape=[6, 256, feature_dims[0], feature_dims[1]],
+                    max_shape=[6, 256, feature_dims[0], feature_dims[1]],
+                ),
+            )
+        )
+    ],
+)
+
+onnx_config = dict(
+    type="onnx",
+    export_params=True,
+    keep_initializers_as_inputs=False,
+    opset_version=17,
+    save_file="camera_point_bev.onnx",
+    input_names=["points", "lidar2image", "img_aug_matrix", "geom_feats", "kept", "ranks", "indices", "image_feats"],
+    output_names=["bbox_pred", "score", "label_pred"],
+    dynamic_axes={
+        "points": {
+            0: "num_points",
+        },
+        "lidar2image": {
+            0: "num_imgs",
+        },
+        "img_aug_matrix": {
+            0: "num_imgs",
+        },
+        "geom_feats": {
+            0: "num_kept",
+        },
+        "kept": {
+            0: "num_geom_feats",
+        },
+        "ranks": {
+            0: "num_kept",
+        },
+        "indices": {
+            0: "num_kept",
+        },
+        "image_feats": {
+            0: "num_imgs",
+        },
+    },
+    input_shape=None,
+    verbose=True,
+)
@@ -0,0 +1,49 @@
+codebase_config = dict(type="mmdet3d", task="VoxelDetection", model_type="end2end")
+
+custom_imports = dict(
+    imports=[
+        "projects.BEVFusion.deploy",
+        "projects.BEVFusion.bevfusion",
+        "projects.SparseConvolution",
+    ],
+    allow_failed_imports=False,
+)
+
+backend_config = dict(
+    type="tensorrt",
+    common_config=dict(max_workspace_size=1 << 32),
+    model_inputs=[
+        dict(
+            input_shapes=dict(
+                voxels=dict(
+                    min_shape=[1, 10, 5], opt_shape=[64000, 10, 5], max_shape=[256000, 10, 5]
+                ),  # [M, maximum number of points, features] features=5 when using intensity
+                coors=dict(min_shape=[1, 3], opt_shape=[64000, 3], max_shape=[256000, 3]),
+                num_points_per_voxel=dict(min_shape=[1], opt_shape=[64000], max_shape=[256000]),
+            )
+        )
+    ],
+)
+
+onnx_config = dict(
+    type="onnx",
+    export_params=True,
+    keep_initializers_as_inputs=False,
+    opset_version=17,
+    save_file="main_body.onnx",
+    input_names=["voxels", "coors", "num_points_per_voxel"],
+    output_names=["bbox_pred", "score", "label_pred"],
+    dynamic_axes={
+        "voxels": {
+            0: "voxels_num",
+        },
+        "coors": {
+            0: "voxels_num",
+        },
+        "num_points_per_voxel": {
+            0: "voxels_num",
+        },
+    },
+    input_shape=None,
+    verbose=True,
+)
@@ -1,7 +1,3 @@
-from . import base
 from .voxel_detection import VoxelDetection
 
-__all__ = [
-    "base",
-    "VoxelDetection",
-]
+__all__ = ["VoxelDetection"]