feat(rgbd): add RGB-D reconstruction pipeline

Add end-to-end RGB-D reconstruction support across the C++ core and Python API.

- add a native merge_rgbd_views path, view-aware 3D pose containers, and nanobind bindings

- expose Python helpers to sample aligned depth, apply per-joint offsets, lift UVD poses to world space, and run reconstruct_rgbd

- add RGB-D regression tests for merging, manual pipeline parity, symmetric depth sampling windows, and out-of-bounds joints

- bump the project version from 0.1.0 to 0.2.0 for the new feature surface
This commit is contained in:
2026-03-26 13:04:57 +08:00
parent 6c09f7044b
commit ed721729fd
12 changed files with 1932 additions and 5 deletions
+1
View File
@@ -3,6 +3,7 @@
set(RPT_SOURCES
camera.cpp
interface.cpp
rgbd_merger.cpp
triangulator.cpp
)
+37
View File
@@ -23,6 +23,17 @@ size_t pose3d_offset(size_t person, size_t joint, size_t coord, size_t num_joint
{
return (((person * num_joints) + joint) * 4) + coord;
}
size_t pose3d_by_view_offset(
size_t view,
size_t person,
size_t joint,
size_t coord,
size_t max_persons,
size_t num_joints)
{
return ((((view * max_persons) + person) * num_joints) + joint) * 4 + coord;
}
} // namespace
// =================================================================================================
@@ -53,6 +64,11 @@ const float &TrackedPoseBatch3DView::at(size_t person, size_t joint, size_t coor
return data[pose3d_offset(person, joint, coord, num_joints)];
}
const float &PoseBatch3DByViewView::at(size_t view, size_t person, size_t joint, size_t coord) const
{
return data[pose3d_by_view_offset(view, person, joint, coord, max_persons, num_joints)];
}
const float &PoseBatch2D::at(size_t view, size_t person, size_t joint, size_t coord) const
{
return data[pose2d_offset(view, person, joint, coord, max_persons, num_joints)];
@@ -129,6 +145,27 @@ PoseBatch3DView PoseBatch3D::view() const
return PoseBatch3DView {data.data(), num_persons, num_joints};
}
float &PoseBatch3DByView::at(size_t view, size_t person, size_t joint, size_t coord)
{
return data[pose3d_by_view_offset(view, person, joint, coord, max_persons, num_joints)];
}
const float &PoseBatch3DByView::at(size_t view, size_t person, size_t joint, size_t coord) const
{
return data[pose3d_by_view_offset(view, person, joint, coord, max_persons, num_joints)];
}
PoseBatch3DByViewView PoseBatch3DByView::view() const
{
return PoseBatch3DByViewView {
data.data(),
person_counts.data(),
num_views,
max_persons,
num_joints,
};
}
NestedPoses3D PoseBatch3D::to_nested() const
{
NestedPoses3D poses_3d(num_persons);
+37
View File
@@ -45,6 +45,17 @@ struct TrackedPoseBatch3DView
const float &at(size_t person, size_t joint, size_t coord) const;
};
struct PoseBatch3DByViewView
{
const float *data = nullptr;
const uint32_t *person_counts = nullptr;
size_t num_views = 0;
size_t max_persons = 0;
size_t num_joints = 0;
const float &at(size_t view, size_t person, size_t joint, size_t coord) const;
};
struct PoseBatch2D
{
std::vector<float> data;
@@ -74,6 +85,19 @@ struct PoseBatch3D
static PoseBatch3D from_nested(const NestedPoses3D &poses_3d);
};
struct PoseBatch3DByView
{
std::vector<float> data;
std::vector<uint32_t> person_counts;
size_t num_views = 0;
size_t max_persons = 0;
size_t num_joints = 0;
float &at(size_t view, size_t person, size_t joint, size_t coord);
const float &at(size_t view, size_t person, size_t joint, size_t coord) const;
PoseBatch3DByViewView view() const;
};
// =================================================================================================
struct PairCandidate
@@ -242,6 +266,11 @@ PoseBatch3D triangulate_poses(
const TriangulationConfig &config,
const TriangulationOptions *options_override = nullptr);
PoseBatch3D merge_rgbd_views(
const PoseBatch3DByViewView &poses_3d,
const TriangulationConfig &config,
float max_distance = 0.5f);
TriangulationResult triangulate_with_report(
const PoseBatch2DView &poses_2d,
const TriangulationConfig &config,
@@ -256,6 +285,14 @@ inline PoseBatch3D triangulate_poses(
return triangulate_poses(poses_2d.view(), config, options_override);
}
inline PoseBatch3D merge_rgbd_views(
const PoseBatch3DByView &poses_3d,
const TriangulationConfig &config,
float max_distance = 0.5f)
{
return merge_rgbd_views(poses_3d.view(), config, max_distance);
}
inline TriangulationTrace triangulate_debug(
const PoseBatch2D &poses_2d,
const TriangulationConfig &config,
+1327
View File
File diff suppressed because it is too large Load Diff