reusex::vision Namespace Reference

Namespaces
namespace	common
namespace	libtorch
namespace	onnx
namespace	osd
namespace	tensor_rt
namespace	torchvision

Classes
struct	AnnotationConfig
	Configuration for ML annotation inference and data loading. More...
class	BackendFactory
class	Dataloader
struct	IData
class	IDataset
class	IMLBackend
class	IModel
class	IVideoModel
struct	SegmentPanoramaOptions
	Options controlling how a panorama is tiled and segmented. More...

Enumerations
enum class	Backend {   opencv , tensor_rt , libtorch , dnn ,   onnx_runtime , openvino , unknown }
enum class	Model { yolo , sam3 , sam3p1 }

Functions
auto	annotate (const std::filesystem::path &dbPath, const std::filesystem::path &modelPath, const AnnotationConfig &config=AnnotationConfig{}) -> int
	Run semantic annotation on sensor frames using ML models.
bool	is_sequence_boundary (std::size_t index, int node_id, int prev_id)
	Whether the video tracker must reset before the frame at index.
std::unique_ptr< IModel >	create_model_from_path (const std::filesystem::path &modelPath, bool use_cuda)
	Detect the backend + model kind from modelPath and build the model.
torch::Tensor	xyxy_to_xywh (const torch::Tensor &x)
	Convert bounding boxes from (x1,y1,x2,y2) to (cx,cy,w,h) format.
torch::Tensor	xywh_to_xyxy (const torch::Tensor &x)
	Convert bounding boxes from (cx,cy,w,h) to (x1,y1,x2,y2) format.
torch::Tensor	nms (const torch::Tensor &bboxes, const torch::Tensor &scores, float iou_threshold=0.45)
	Non-maximum suppression on bounding boxes.
torch::Tensor	non_max_suppression (torch::Tensor predictions, float confThreshold=0.25, float iouThreshold=0.45, int maxDetections=300)
	YOLO-style non-maximum suppression with class-aware filtering.
auto	project (const ProjectDB &db, CloudConstPtr cloud) -> CloudLPtr
	Project the labels stored in the database on tho the assebled point cloud.
cv::Mat	segment_panorama (IModel &model, const cv::Mat &equirect_bgr, const SegmentPanoramaOptions &opts)
	Segment a 360 equirectangular panorama with a SAM3 model.
float	generate_scale (const cv::Mat &image, const cv::Size &target_size, bool scale_up=false)
float	letterbox (cv::Mat &input_image, cv::Mat &output_image, const cv::Size &target_size)
float	cropbox (cv::Mat &input_image, cv::Mat &output_image, const cv::Size &target_size)

Variables
std::vector< std::string >	Yolov8_className
	YOLO v8 class names for object detection.

Enumeration Type Documentation

Backend

enum class reusex::vision::Backend

strong

Enumerator
opencv
tensor_rt
libtorch
dnn
onnx_runtime
openvino
unknown

Definition at line 32 of file BackendFactory.hpp.

Model

enum class reusex::vision::Model

strong

Enumerator
yolo
sam3
sam3p1

Definition at line 15 of file IMLBackend.hpp.

Function Documentation

annotate()

auto reusex::vision::annotate	(	const std::filesystem::path &	dbPath,
		const std::filesystem::path &	modelPath,
		const AnnotationConfig &	config = AnnotationConfig{} ) -> int

Run semantic annotation on sensor frames using ML models.

Parameters

dbPath	Path to project database containing sensor frames
modelPath	Path to ML model (.pt, .engine, .onnx)
config	Configuration for inference and data loading

Returns

0 on success, error code on failure

create_model_from_path()

std::unique_ptr< IModel > reusex::vision::create_model_from_path	(	const std::filesystem::path &	modelPath,
		bool	use_cuda )

Detect the backend + model kind from modelPath and build the model.

Exceptions

std::runtime_error

if the required backend is not compiled in.

cropbox()

float reusex::vision::cropbox	(	cv::Mat &	input_image,
		cv::Mat &	output_image,
		const cv::Size &	target_size )

generate_scale()

float reusex::vision::generate_scale	(	const cv::Mat &	image,
		const cv::Size &	target_size,
		bool	scale_up = false )

is_sequence_boundary()

bool reusex::vision::is_sequence_boundary ( std::size_t index, int node_id, int prev_id )

inline

Whether the video tracker must reset before the frame at index.

A stateful video model must start a fresh temporal memory at the beginning of every ordered sequence. Within one scan the sensor-frame node ids increase monotonically (small forward gaps from dropped frames are normal and must NOT trigger a reset), so a boundary is only the very first frame (index 0) or a point where the node id fails to increase — which happens when a dataset concatenates multiple sequences that restart their numbering.

Kept as a free function (independent of any tracker/dataset object) so the boundary logic can be unit-tested directly.

Parameters

index	Position in the (possibly prefix-trimmed) ordered list.
node_id	Node id of the frame at index.
prev_id	Node id of the frame at index-1 (unused when index==0).

Returns

true if the tracker should reset before processing this frame.

Definition at line 63 of file annotate.hpp.

letterbox()

float reusex::vision::letterbox	(	cv::Mat &	input_image,
		cv::Mat &	output_image,
		const cv::Size &	target_size )

nms()

torch::Tensor reusex::vision::nms	(	const torch::Tensor &	bboxes,
		const torch::Tensor &	scores,
		float	iou_threshold = 0.45 )

Non-maximum suppression on bounding boxes.

The algorithm is torchvision's reference CPU kernel, vendored under BSD-3-Clause in vision/third_party/torchvision_nms_kernel.hpp (#141). This wrapper owns the ReUseX-facing contract: CPU, float32, empty input tolerated.

Boxes are suppressed when IoU is strictly greater than iou_threshold, so a threshold of 1.0 suppresses nothing and ties at the threshold survive. Equal scores are broken towards the lower index (the sort is stable).

Parameters

bboxes	Bounding boxes tensor [N, 4] in xyxy format. CPU, float32.
scores	Confidence scores tensor [N]. CPU, float32.
iou_threshold	IoU threshold for suppression.

Returns

Indices of kept boxes, in descending score order (int64).

Exceptions

c10::Error

if the tensors are not CPU float32, or are malformed.

non_max_suppression()

torch::Tensor reusex::vision::non_max_suppression	(	torch::Tensor	predictions,
		float	confThreshold = 0.25,
		float	iouThreshold = 0.45,
		int	maxDetections = 300 )

YOLO-style non-maximum suppression with class-aware filtering.

Parameters

predictions	Raw model output [batch_size, 116, 8400].
confThreshold	Confidence threshold for filtering.
iouThreshold	IoU threshold for NMS.
maxDetections	Maximum detections to keep per image.

Returns

Filtered detections [batch_size, maxDetections, 6+32].

project()

auto reusex::vision::project	(	const ProjectDB &	db,
		CloudConstPtr	cloud ) -> CloudLPtr

Project the labels stored in the database on tho the assebled point cloud.

For each view in the database this constructs a croped point cloud, computes a z-buffer and the assigns the precomputed labels to the closses point in the point cloud.

segment_panorama()

cv::Mat reusex::vision::segment_panorama	(	IModel &	model,
		const cv::Mat &	equirect_bgr,
		const SegmentPanoramaOptions &	opts )

Segment a 360 equirectangular panorama with a SAM3 model.

Parameters

model	SAM3 model handle (created via BackendFactory).
equirect_bgr	Equirectangular panorama (BGR, CV_8UC3).
opts	Tiling / prompt / confidence options.

Returns

A CV_32S equirect label map the same size as equirect_bgr, with -1 for background. Class ids are stable across tiles because a single model instance assigns each prompt text a fixed class id.

xywh_to_xyxy()

torch::Tensor reusex::vision::xywh_to_xyxy

(

const torch::Tensor &

x

)

Convert bounding boxes from (cx,cy,w,h) to (x1,y1,x2,y2) format.

xyxy_to_xywh()

torch::Tensor reusex::vision::xyxy_to_xywh

(

const torch::Tensor &

x

)

Convert bounding boxes from (x1,y1,x2,y2) to (cx,cy,w,h) format.

Variable Documentation

Yolov8_className

std::vector<std::string> reusex::vision::Yolov8_className

extern

YOLO v8 class names for object detection.

Contains the list of class names that can be detected by YOLO v8 model.