Add depth_to_point_cloud operator and demo

applications/CMakeLists.txt

@@ -51,6 +51,8 @@ add_holohub_application(deltacast_receiver DEPENDS
 
 add_holohub_application(depth_anything_v2)
 
+add_holohub_application(depth_to_point_cloud_demo DEPENDS OPERATORS depth_to_point_cloud)
+
 add_subdirectory(distributed)
 
 add_holohub_application(endoscopy_depth_estimation)

applications/depth_to_point_cloud_demo/CMakeLists.txt

@@ -0,0 +1,33 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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.
+
+project(depth_to_point_cloud_demo NONE)
+
+find_package(holoscan 4.0.0 REQUIRED CONFIG
+             PATHS "/opt/nvidia/holoscan" "/workspace/holoscan-sdk/install")
+
+if(BUILD_TESTING)
+  add_test(NAME depth_to_point_cloud_demo_python_test
+    COMMAND python3 ${CMAKE_CURRENT_SOURCE_DIR}/depth_to_point_cloud_demo.py --frames 10
+    WORKING_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+  set_property(TEST depth_to_point_cloud_demo_python_test PROPERTY ENVIRONMENT
+               "PYTHONPATH=${GXF_LIB_DIR}/../python/lib:${CMAKE_BINARY_DIR}/python/lib")
+
+  set_tests_properties(depth_to_point_cloud_demo_python_test
+    PROPERTIES
+      PASS_REGULAR_EXPRESSION "valid="
+      FAIL_REGULAR_EXPRESSION "(^|[^a-z])Error;ERROR;Failed")
+endif()

applications/depth_to_point_cloud_demo/README.md

@@ -0,0 +1,63 @@
+# Depth to Point Cloud Demo
+
+A minimal, hardware-free demo of the [`depth_to_point_cloud`](../../operators/depth_to_point_cloud)
+operator. It generates a synthetic organized depth image (a gently tilting plane) plus an aligned
+RGB image entirely on the GPU, deprojects it into an organized `H x W x 3` point cloud, and reports
+the valid-point count and Z range each frame. No camera or dataset is required.
+
+## Run
+
+```bash
+./holohub run depth_to_point_cloud_demo
+# or directly:
+python3 applications/depth_to_point_cloud_demo/depth_to_point_cloud_demo.py --frames 100
+```
+
+Expected output (per frame):
+
+```text
+[depth_to_point_cloud_demo] points=307200 valid=307200 z=[1.xxx, 2.xxx] m
+```
+
+## Pipeline
+
+```text
+SyntheticDepthGeneratorOp  --depth-->  DepthToPointCloudOp  --point_cloud-->  PointCloudStatsOp
+                           --color-->
+```
+
+## Using a real Intel RealSense camera
+
+Replace the synthetic generator with the [`realsense_camera`](../../operators/realsense_camera)
+operator. It applies librealsense's `units_transform` internally, so `depth_buffer` is emitted as
+**`GRAY32F` float32 already in meters** (in a `VideoBuffer`), and `color_buffer` as `RGBA8`. Because
+the depth is metric, use `depth_scale=1.0` — **not** `0.001`; the `0.001` (uint16 millimeters) value
+is only for raw `Z16` sources that have not been unit-transformed:
+
+```python
+from holohub.realsense_camera import RealsenseCameraOp
+
+camera = RealsenseCameraOp(self, name="camera")
+cloud = DepthToPointCloudOp(
+    self, name="point_cloud", allocator=...,
+    fx=fx, fy=fy, cx=cx, cy=cy,   # from the camera's depth_camera_model intrinsics
+    depth_scale=1.0,              # depth_buffer is float32 meters (units_transform applied)
+    depth_min=0.1, depth_max=10.0,
+)
+self.add_flow(camera, cloud, {("depth_buffer", "depth")})
+self.add_flow(camera, cloud, {("color_buffer", "color")})  # color_channels = 4 (RGBA)
+```
+
+The camera also exposes `depth_camera_model` / `color_camera_model` outputs carrying the per-stream
+intrinsics, which can drive the optional `intrinsics` input instead of static `fx/fy/cx/cy`.
+
+## Interactive 3D visualization
+
+To view the cloud, replace `PointCloudStatsOp` with `HolovizOp` configured to render the
+`point_cloud` tensor as 3D points (and the `colors` tensor for per-point color). Reshaping the
+organized `H x W x 3` cloud to `N x 3` and dropping invalid (NaN) points first is recommended.
+
+## Requirements
+
+- Holoscan SDK ≥ 4.0.0, CUDA, CuPy. Builds the `depth_to_point_cloud` operator (declared as a
+  dependency). Platforms: `x86_64`, `aarch64`.

applications/depth_to_point_cloud_demo/depth_to_point_cloud_demo.py

@@ -0,0 +1,176 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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.
+
+"""Demo for DepthToPointCloudOp.
+
+Generates a synthetic organized depth image (a gently tilting plane) plus an aligned RGB
+image entirely on the GPU, deprojects it into an organized point cloud with
+DepthToPointCloudOp, and validates the result. No camera or dataset is required, so the
+app runs in CI. See README.md for wiring a real Intel RealSense camera or adding Holoviz
+3D rendering.
+"""
+
+import argparse
+
+import cupy as cp
+from holoscan.conditions import CountCondition
+from holoscan.core import Application, Operator, OperatorSpec
+from holoscan.resources import BlockMemoryPool, CudaStreamPool, MemoryStorageType
+
+from holohub.depth_to_point_cloud import DepthToPointCloudOp
+
+
+class SyntheticDepthGeneratorOp(Operator):
+    """Emit a synthetic float32 depth image (meters) and an aligned uint8 RGB image."""
+
+    def __init__(self, fragment, *args, width=640, height=480, **kwargs):
+        self.width = width
+        self.height = height
+        self.frame = 0
+        ys, xs = cp.meshgrid(
+            cp.arange(height, dtype=cp.float32),
+            cp.arange(width, dtype=cp.float32),
+            indexing="ij",
+        )
+        self._xs = xs
+        self._ys = ys
+        super().__init__(fragment, *args, **kwargs)
+
+    def setup(self, spec: OperatorSpec):
+        spec.output("depth")
+        spec.output("color")
+
+    def compute(self, op_input, op_output, context):
+        t = self.frame * 0.05
+        # A tilted plane in meters: ~1.0 m near the top-left, increasing across the frame,
+        # with a slow global oscillation so successive frames differ.
+        depth = (
+            1.0 + 0.5 * (self._xs / self.width) + 0.4 * (self._ys / self.height) + 0.3 * cp.sin(t)
+        ).astype(cp.float32)
+
+        r = (255.0 * self._xs / self.width).astype(cp.uint8)
+        g = (255.0 * self._ys / self.height).astype(cp.uint8)
+        b = cp.full_like(r, 128)
+        color = cp.ascontiguousarray(cp.stack([r, g, b], axis=-1))  # HxWx3 uint8
+
+        op_output.emit({"depth": depth}, "depth")
+        op_output.emit({"color": color}, "color")
+        self.frame += 1
+
+
+class PointCloudStatsOp(Operator):
+    """Pull the point cloud and report valid-point count and Z range (CI-friendly sink)."""
+
+    def setup(self, spec: OperatorSpec):
+        spec.input("in")
+
+    def compute(self, op_input, op_output, context):
+        msg = op_input.receive("in")
+        pc = cp.asarray(msg["point_cloud"])  # HxWx3 float32
+        z = pc[..., 2]
+        valid = ~cp.isnan(z)
+        n_valid = int(valid.sum().get())
+
+        # The colored path is connected, so a colors tensor must accompany the cloud and
+        # share its H x W footprint (3 uint8 channels).
+        colors = cp.asarray(msg["colors"])  # HxWx3 uint8
+        if colors.shape[:2] != pc.shape[:2] or colors.shape[2] != 3:
+            raise RuntimeError(
+                f"colors shape {colors.shape} does not match cloud {pc.shape[:2]} x 3"
+            )
+
+        if n_valid:
+            zmin = float(z[valid].min().get())
+            zmax = float(z[valid].max().get())
+            print(
+                f"[depth_to_point_cloud_demo] points={pc.shape[0] * pc.shape[1]} "
+                f"valid={n_valid} z=[{zmin:.3f}, {zmax:.3f}] m colors={tuple(colors.shape)}"
+            )
+        else:
+            print("[depth_to_point_cloud_demo] no valid points")
+
+
+class DepthToPointCloudDemoApp(Application):
+    def __init__(self, frames=100, width=640, height=480):
+        super().__init__()
+        self._frames = frames
+        self._width = width
+        self._height = height
+
+    def compose(self):
+        generator = SyntheticDepthGeneratorOp(
+            self,
+            CountCondition(self, count=self._frames),
+            name="generator",
+            width=self._width,
+            height=self._height,
+        )
+
+        # Two device tensors per frame (HxWx3 float32 point cloud + HxWx3 uint8 colors) drawn
+        # from this pool; size each block for the larger (float32 XYZ) output and keep enough
+        # blocks for both tensors plus one frame of pipelining headroom.
+        out_blocks = 4
+        block_size = self._width * self._height * 3 * 4  # float32 XYZ is the larger output
+        cloud = DepthToPointCloudOp(
+            self,
+            name="point_cloud",
+            allocator=BlockMemoryPool(
+                self,
+                name="pool",
+                storage_type=MemoryStorageType.DEVICE,
+                block_size=block_size,
+                num_blocks=out_blocks,
+            ),
+            # Pinhole intrinsics for the synthetic camera: square pixels (fx == fy) with the
+            # principal point at the image center. A single focal length is used for both axes
+            # by design; the focal length is independent of the image aspect ratio.
+            fx=float(self._width) * 0.8,
+            fy=float(self._width) * 0.8,
+            cx=(self._width - 1) / 2.0,
+            cy=(self._height - 1) / 2.0,
+            depth_scale=1.0,  # synthetic depth is already in meters
+            depth_min=0.1,
+            depth_max=10.0,
+            cuda_stream_pool=CudaStreamPool(self, name="stream_pool", reserved_size=4),
+        )
+
+        sink = PointCloudStatsOp(self, name="stats")
+
+        self.add_flow(generator, cloud, {("depth", "depth")})
+        self.add_flow(generator, cloud, {("color", "color")})
+        self.add_flow(cloud, sink, {("point_cloud", "in")})
+
+
+def main():
+    parser = argparse.ArgumentParser(description="DepthToPointCloudOp synthetic demo")
+    parser.add_argument("--frames", type=int, default=100, help="Number of frames to process")
+    parser.add_argument("--width", type=int, default=640)
+    parser.add_argument("--height", type=int, default=480)
+    args = parser.parse_args()
+
+    for name, value in (
+        ("--frames", args.frames),
+        ("--width", args.width),
+        ("--height", args.height),
+    ):
+        if value <= 0:
+            parser.error(f"{name} must be a positive integer (got {value})")
+
+    app = DepthToPointCloudDemoApp(frames=args.frames, width=args.width, height=args.height)
+    app.run()
+
+
+if __name__ == "__main__":
+    main()

applications/depth_to_point_cloud_demo/metadata.json

@@ -0,0 +1,42 @@
+{
+	"$schema": "urn:holohub:application:v1",
+	"application": {
+		"name": "Depth to Point Cloud Demo",
+		"description": "Hardware-free demo of DepthToPointCloudOp: a synthetic GPU depth+RGB generator is deprojected into an organized point cloud and validated each frame.",
+		"authors": [
+			{
+				"name": "Nitheesh Kumar",
+				"affiliation": "Zobot Lab"
+			}
+		],
+		"language": "Python",
+		"version": "0.1.0",
+		"changelog": {
+			"0.1.0": "Initial release: synthetic depth generator -> DepthToPointCloudOp -> validation sink."
+		},
+		"holoscan_sdk": {
+			"minimum_required_version": "4.0.0",
+			"tested_versions": [
+				"4.3.0"
+			]
+		},
+		"platforms": [
+			"x86_64",
+			"aarch64"
+		],
+		"tags": ["Computer Vision and Perception", "Robotics", "Point Cloud", "Depth", "3D"],
+		"ranking": 3,
+		"requirements": {
+			"python-packages": [
+				{
+					"name": "cupy",
+					"version": "13.6.0"
+				}
+			]
+		},
+		"run": {
+			"command": "python3 <holohub_app_source>/depth_to_point_cloud_demo.py",
+			"workdir": "holohub_bin"
+		}
+	}
+}

operators/CMakeLists.txt

@@ -24,6 +24,7 @@ add_holohub_operator(cvcuda_holoscan_interop)
 add_subdirectory(deidentification)
 add_subdirectory(dds)
 add_holohub_operator(deltacast_videomaster)
+add_holohub_operator(depth_to_point_cloud)
 add_holohub_operator(display_gpu_resident)
 add_holohub_operator(ucxx_send_receive)
 add_holohub_operator(emergent_source DEPENDS EXTENSIONS emergent_source)

operators/depth_to_point_cloud/CMakeLists.txt

@@ -0,0 +1,54 @@
+# SPDX-FileCopyrightText: Copyright (c) 2026 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# 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.
+
+cmake_minimum_required(VERSION 3.24)
+
+project(depth_to_point_cloud LANGUAGES CXX CUDA)
+
+find_package(holoscan 4.0.0 REQUIRED CONFIG
+             PATHS "/opt/nvidia/holoscan" "/workspace/holoscan-sdk/install")
+
+add_library(depth_to_point_cloud SHARED
+  depth_to_point_cloud.cpp
+  depth_to_point_cloud.hpp
+  deproject.cu
+  deproject.hpp
+)
+
+set_target_properties(depth_to_point_cloud
+  PROPERTIES
+    # compile for the architecture of the current GPU
+    CUDA_ARCHITECTURES "native"
+)
+
+target_link_libraries(depth_to_point_cloud
+  PUBLIC
+    holoscan::core
+)
+
+target_include_directories(depth_to_point_cloud
+  INTERFACE
+    ${CMAKE_CURRENT_SOURCE_DIR}
+)
+
+if(HOLOHUB_BUILD_PYTHON)
+    add_subdirectory(python)
+endif()
+
+if(BUILD_TESTING)
+    add_subdirectory(test)
+endif()
+
+install(TARGETS depth_to_point_cloud)