Selection and preprocessing knobs for building the training set. More...

#include <TrainingViews.hpp>

Public Attributes
bool	include_frames = true
bool	include_panorama_slices = false
	Also dice content-aligned 360 panoramas into overlapping tangent views.
int	pano_n_yaw = 8
	equator slices per panorama
double	pano_fov_deg = 90.0
	per-slice horizontal FOV [deg]
int	pano_tile = 1024
	slice size [px, square]
int	frame_stride = 1
	Take every Nth sensor frame.
int	first_frame = -1
	Inclusive sensor-frame node-id range, -1 for unbounded.
int	last_frame = -1
int	max_image_size = 0
	Downscale images (and intrinsics) so the long edge is at most this many pixels.
std::size_t	max_views = 0
	Hard cap on the number of views (0 = unlimited), applied after striding.

Detailed Description

Selection and preprocessing knobs for building the training set.

Definition at line 39 of file TrainingViews.hpp.

Member Data Documentation

◆ first_frame

int reusex::gsplat::TrainingViewOptions::first_frame = -1

Inclusive sensor-frame node-id range, -1 for unbounded.

Validation on NewOffice showed a building-wide sparse sample renders black while a contiguous capture segment reconstructs cleanly, so region selection is a first-class knob rather than an afterthought (docs/research/gaussian-splatting.md).

Definition at line 60 of file TrainingViews.hpp.

◆ frame_stride

int reusex::gsplat::TrainingViewOptions::frame_stride = 1

Take every Nth sensor frame.

3DGS wants dense multi-view overlap, so the default keeps every frame; raise it to trade coverage for speed.

Definition at line 53 of file TrainingViews.hpp.

◆ include_frames

bool reusex::gsplat::TrainingViewOptions::include_frames = true

Definition at line 40 of file TrainingViews.hpp.

◆ include_panorama_slices

bool reusex::gsplat::TrainingViewOptions::include_panorama_slices = false

Also dice content-aligned 360 panoramas into overlapping tangent views.

Only panoramas whose pose_source == "aligned" are used — a panorama carrying only its timestamp-matched neighbour pose is not trustworthy enough to train against.

Definition at line 46 of file TrainingViews.hpp.

◆ last_frame

int reusex::gsplat::TrainingViewOptions::last_frame = -1

Definition at line 61 of file TrainingViews.hpp.

◆ max_image_size

int reusex::gsplat::TrainingViewOptions::max_image_size = 0

Downscale images (and intrinsics) so the long edge is at most this many pixels.

0 disables resizing. Training cost is per-pixel, so this is the single most effective speed knob.

Definition at line 66 of file TrainingViews.hpp.

◆ max_views

std::size_t reusex::gsplat::TrainingViewOptions::max_views = 0

Hard cap on the number of views (0 = unlimited), applied after striding.

Definition at line 69 of file TrainingViews.hpp.

◆ pano_fov_deg

double reusex::gsplat::TrainingViewOptions::pano_fov_deg = 90.0

per-slice horizontal FOV [deg]

Definition at line 48 of file TrainingViews.hpp.

◆ pano_n_yaw

int reusex::gsplat::TrainingViewOptions::pano_n_yaw = 8

equator slices per panorama

Definition at line 47 of file TrainingViews.hpp.

◆ pano_tile

int reusex::gsplat::TrainingViewOptions::pano_tile = 1024

slice size [px, square]

Definition at line 49 of file TrainingViews.hpp.

The documentation for this struct was generated from the following file:

libs/reusex/include/gsplat/TrainingViews.hpp

Public Attributes