Depth image renderer

[victorreijgwart]

Description

TODO: Thank you for opening a PR! Please summarize the changes in 1–2 sentences.

Thanks a lot to @Divelix for suggesting and implementing this feature!

This PR should only be reviewed and merged after PR #88.

Type of change

• Bug fix (non-breaking change which fixes an issue)
• New feature (non-breaking change which adds functionality)
• Breaking change (fix or feature that causes existing functionality to not work as expected)
• Other (please describe):

Detailed Summary

TODO: Provide the motivation, context, and links to any related issues, PRs, or documentation:

• Motivation: Why is this change necessary?
• Context: How does it fit into wavemap's functionality?
• Related issues/PRs: Fixes # (issue) / Links to other PRs

API Changes

TODO: List any changes to wavemap's APIs to help users update their code. Write "None" if there are no changes.

C++ API:

Python API:

ROS1 Interface:

Review Notes

TODO: Is there anything specific the reviewers should focus on, or are there unresolved questions? Mention them here.

Testing

Automated Tests

TODO: Have you added or modified unit tests to verify these changes? If not, let us know if you'd like assistance.

Manual Tests

TODO: If manual tests were performed to verify these changes, describe them here and include instructions to reproduce them.
Describe test configurations where applicable.

System information (optional):

• CPU: [e.g., Intel i9-9900K]
• GPU: [e.g., Nvidia RTX 2080Ti]
• RAM: [e.g., 32GB]
• OS: [e.g., Ubuntu 20.04]
• API: [e.g., C++, Python, ROS1]
• Installation: [e.g., pre-built Docker, local CMake, Pip, catkin]

Runtime information (optional):

• Launch file: [e.g., Link or GitHub Gist]
• Config file: [e.g., Link or GitHub Gist]
• Dataset name (if public): [e.g., Newer College Cloister]
• Custom setup (for private datasets, or live usage):
  • Depth sensor: [e.g., Livox MID360 LiDAR]
  • Pose source: [e.g., Odometry from FastLIO2]

For performance or accuracy-related changes, include the above system and runtime information and describe:

• Performance (optional)
  • Measured operation: [e.g. serializing the map, performing 1M queries, processing dataset X]
  • Metrics [e.g., CPU time, wall time, total RAM usage]
• Accuracy (optional)
  • Metrics: [e.g., AUC, accuracy, recall]
• Summary of changes
  • What metrics improved and by how much?
  • Did any metrics worsen?

Benchmarks (To be completed by maintainers)

TODO: We will rerun wavemap's benchmarks and report the results here to validate there are no general performance/accuracy regressions.

Checklist

General

• My code follows the style guidelines of this project
• I have performed a self-review of my code
• I have commented my code, particularly in hard-to-understand areas
• I have added or updated tests as required
• Any required changes in dependencies have been committed and pushed

Documentation (where applicable)

• I have updated the installation instructions (in docs/pages/installation)
• I have updated the code's inline API documentation (e.g., docstrings)
• I have updated the parameter documentation (in docs/pages/parameters)
• I have updated/extended the tutorials (in docs/pages/tutorials)

[JonasFrey96]

Hi Victor,

I ran into some troubles using the raycasting:
Here is a minimal script explaining what I tried to do:

import wavemap as wm
import numpay as np

W, H = 1920, 1280


# rot_so3, trans, K = ...

your_map = wm.Map.load("path/to/your_map.wvmp")
renderer = wm.RaycastingRenderer(
    your_map,
    wm.Projector.create(
        {
            "type": "pinhole_camera_projector",
            "width": W,
            "height": H,
            "fx": K[0, 0],
            "fy": K[1, 1],
            "cx": K[0, 2],
            "cy": K[1, 2],
        }
    ),
    0.1,
    150,
    -1,
)

pose = wm.Pose(wm.Rotation(rot_so3), trans)
undistorted_depth_image = renderer.render(pose).data


import open3d as o3d  # version 0.19.0

depth = o3d.t.geometry.Image(o3d.core.Tensor(undistorted_depth_image[::-1, ::], dtype=o3d.core.Dtype.Float32))
intrinsic = o3d.camera.PinholeCameraIntrinsic()
intrinsic.set_intrinsics(
    width=W,
    height=H,
    fx=K[0, 0],
    fy=K[1, 1],
    cx=K[0, 2],
    cy=K[1, 0],
)
intrinsic_matrix = o3d.core.Tensor(intrinsic.intrinsic_matrix, o3d.core.Dtype.Float64)
extrinsic = np.eye(4)
extrinsic[:3, :3] = rot_so3
extrinsic[:3, 3] = trans

pcd = o3d.t.geometry.PointCloud.create_from_depth_image(
    depth,
    intrinsic_matrix,
    o3d.core.Tensor(extrinsic, dtype=o3d.core.Dtype.Float64),
    depth_scale=1,
    depth_max=150,
).to_legacy()

# Visualize the point cloud
o3d.visualization.draw_geometries([pcd])

I used the following Wavemap

Target Image to Create Depthmap (sets K,H,W,rot,trans)

Projection Points on Image is perfect

The metric depth looked in this viewpoint slightly off.
Camera View:

I found that the depth is slightly distorted. specifically here on the right of the scene the ground should be fully flat but bends a lot.

Therefore, I tried to render an additional viewpoints (one exactly aligning with the camera and one 2m above the camera facing down on the ground).

3D-Person View - Ground

Sideview - Ground

Here we can see a clear distortion of the returned depth values. Any idea how this can happen?

Some further checks that I did:

• I tried using matlab to project the pointcloud (same result so not an open3d error)
• Moving camera further away from the ground it "gets worse"
  
  

[vreijgwart-bdai]

Hi @JonasFrey96,

Thanks for reporting this issue and providing a detailed description!

The bug occurs because we always return a range image, but for depth cameras, we need a depth image. The code to generate both is almost the same, but for depth images we should return the endpoint’s Z-coordinate instead of the Euclidean distance along the ray. At the image center, these values are identical, but they diverge toward the edges -- like a parabola.

Could you please test changing lines 41–46 in raycasting_renderer.cc to directly use the lower-level raycast::first_collision_index function, computing depth in sensor frame as follows?

const auto colliding_index = first_collision_index<MapT>(
    map, W_start_point, W_end_point, log_odds_occupancy_threshold);
if (colliding_index) {
  const FloatingPoint min_cell_width = map.getMinCellWidth();
  const Point3D W_voxel_center =
      convert::indexToCenterPoint(colliding_index.value(), min_cell_width);
  const Point3D C_voxel_center = T_W_C.inverse() * W_voxel_center;
  depth_pixel = projection_model_->cartesianToSensorZ(C_voxel_center);
}

[JonasFrey96]

Hi Victor,
Big thanks this fixed everything:

Here is the correct code snippet that runs:

    const auto colliding_index = raycast::first_collision_index<const MapT>(
        map, W_start_point, W_end_point, log_odds_occupancy_threshold_);
    if (colliding_index) {
      const FloatingPoint min_cell_width = map.getMinCellWidth();
      const Point3D W_voxel_center =
          convert::indexToCenterPoint(colliding_index.value(), min_cell_width);
      const Point3D C_voxel_center = T_W_C.inverse() * W_voxel_center;
      depth_pixel = projection_model_->cartesianToSensorZ(C_voxel_center);
    }