1191 lines
54 KiB
C++
1191 lines
54 KiB
C++
#include "ANSTENSORRTSEG.h"
|
|
#include "Utility.h"
|
|
#include <opencv2/cudaimgproc.hpp>
|
|
#include <future>
|
|
namespace ANSCENTER
|
|
{
|
|
bool TENSORRTSEG::OptimizeModel(bool fp16, std::string& optimizedModelFolder) {
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
if (!ANSODBase::OptimizeModel(fp16, optimizedModelFolder)) {
|
|
return false;
|
|
}
|
|
if (!FileExist(_modelFilePath)) {
|
|
this->_logger.LogFatal("TENSORRTSEG::OptimizeModel", "Raw model file path does not exist", __FILE__, __LINE__);
|
|
return false;
|
|
}
|
|
try {
|
|
_fp16 = fp16;
|
|
optimizedModelFolder = GetParentFolder(_modelFilePath);
|
|
// Check if the engine already exists to avoid reinitializing
|
|
if (!m_trtEngine) {
|
|
// Fixed batch size of 1 for this model
|
|
m_options.optBatchSize = _modelConfig.gpuOptBatchSize;
|
|
m_options.maxBatchSize = _modelConfig.gpuMaxBatchSize;
|
|
m_options.deviceIndex = _modelConfig.gpuDeviceIndex;
|
|
m_options.maxInputHeight = _modelConfig.maxInputHeight;
|
|
m_options.minInputHeight = _modelConfig.minInputHeight;
|
|
m_options.optInputHeight = _modelConfig.optInputHeight;
|
|
m_options.maxInputWidth = _modelConfig.maxInputWidth;
|
|
m_options.minInputWidth = _modelConfig.minInputWidth;
|
|
m_options.optInputWidth = _modelConfig.optInputWidth;
|
|
m_options.engineFileDir = optimizedModelFolder;
|
|
m_options.precision = (_fp16 ? Precision::FP16 : Precision::FP32);
|
|
// Create the TensorRT inference engine
|
|
m_trtEngine = std::make_unique<Engine<float>>(m_options);
|
|
}
|
|
|
|
// Build the TensorRT engine
|
|
auto succ = m_trtEngine->buildWithRetry(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE);
|
|
if (!succ) {
|
|
const std::string errMsg =
|
|
"Error: Unable to build the TensorRT engine. "
|
|
"Try increasing TensorRT log severity to kVERBOSE.";
|
|
this->_logger.LogError("TENSORRTSEG::OptimizeModel", errMsg, __FILE__, __LINE__);
|
|
_modelLoadValid = false;
|
|
return false;
|
|
}
|
|
_modelLoadValid = true;
|
|
return true;
|
|
}
|
|
catch (const std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::OptimizeModel", e.what(), __FILE__, __LINE__);
|
|
optimizedModelFolder.clear();
|
|
return false;
|
|
}
|
|
}
|
|
bool TENSORRTSEG::LoadModel(const std::string& modelZipFilePath, const std::string& modelZipPassword) {
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
try {
|
|
bool result = ANSODBase::LoadModel(modelZipFilePath, modelZipPassword);
|
|
if (!result) return false;
|
|
_modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
|
|
_modelConfig.modelType = ModelType::RTSEG;
|
|
_modelConfig.inpHeight = 640;
|
|
_modelConfig.inpWidth = 640;
|
|
if (_modelConfig.modelMNSThreshold < 0.2)
|
|
_modelConfig.modelMNSThreshold = 0.5;
|
|
if (_modelConfig.modelConfThreshold < 0.2)
|
|
_modelConfig.modelConfThreshold = 0.5;
|
|
if (_modelConfig.numKPS <= 0 || _modelConfig.numKPS > 133) // 133 = COCO wholebody max
|
|
_modelConfig.numKPS = 17;
|
|
if (_modelConfig.kpsThreshold == 0)_modelConfig.kpsThreshold = 0.5; // If not define
|
|
_fp16 = true; // Load Model from Here
|
|
// Load Model from Here
|
|
TOP_K = 100;
|
|
SEG_CHANNELS = 32;
|
|
PROBABILITY_THRESHOLD = _modelConfig.detectionScoreThreshold;
|
|
NMS_THRESHOLD = _modelConfig.modelMNSThreshold;
|
|
SEGMENTATION_THRESHOLD = 0.5f;
|
|
SEG_H = 160;
|
|
SEG_W = 160;
|
|
NUM_KPS = _modelConfig.numKPS;
|
|
KPS_THRESHOLD = _modelConfig.kpsThreshold;
|
|
SEG_CHANNELS = 32; // For segmentation
|
|
|
|
if (!m_trtEngine) {
|
|
// Fixed batch size of 1 for this model
|
|
m_options.optBatchSize = _modelConfig.gpuOptBatchSize;
|
|
m_options.maxBatchSize = _modelConfig.gpuMaxBatchSize;
|
|
m_options.deviceIndex = _modelConfig.gpuDeviceIndex;
|
|
m_options.maxInputHeight = _modelConfig.maxInputHeight;
|
|
m_options.minInputHeight = _modelConfig.minInputHeight;
|
|
m_options.optInputHeight = _modelConfig.optInputHeight;
|
|
m_options.maxInputWidth = _modelConfig.maxInputWidth;
|
|
m_options.minInputWidth = _modelConfig.minInputWidth;
|
|
m_options.optInputWidth = _modelConfig.optInputWidth;
|
|
m_options.engineFileDir = _modelFolder;
|
|
// Use FP16 or FP32 precision based on the input flag
|
|
m_options.precision = (_fp16 ? Precision::FP16 : Precision::FP32);
|
|
// Create the TensorRT inference engine
|
|
m_trtEngine = std::make_unique<Engine<float>>(m_options);
|
|
}
|
|
// 0. Check if the configuration file exist
|
|
if (FileExist(_modelConfigFile)) {
|
|
ModelType modelType;
|
|
std::vector<int> inputShape;
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
|
if (inputShape.size() == 2) {
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
|
}
|
|
}
|
|
else {// This is old version of model zip file
|
|
_modelFilePath = CreateFilePath(_modelFolder, "train_last.onnx");
|
|
_classFilePath = CreateFilePath(_modelFolder, "classes.names");
|
|
std::ifstream isValidFileName(_classFilePath);
|
|
if (!isValidFileName)
|
|
{
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromString();
|
|
}
|
|
else {
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromFile();
|
|
}
|
|
}
|
|
// Load the TensorRT engine file
|
|
if (this->_loadEngineOnCreation) {
|
|
auto succ = m_trtEngine->buildLoadNetwork(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE, m_maxSlotsPerGpu);
|
|
if (!succ) {
|
|
const std::string errMsg = "Error: Unable to load TensorRT engine weights into memory. " + _modelFilePath;
|
|
this->_logger.LogError("TENSORRTSEG::Initialize", errMsg, __FILE__, __LINE__);
|
|
_modelLoadValid = false;
|
|
return false;
|
|
}
|
|
}
|
|
_modelLoadValid = true;
|
|
_isInitialized = true;
|
|
return true;
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::LoadModel", e.what(), __FILE__, __LINE__);
|
|
return false;
|
|
}
|
|
}
|
|
bool TENSORRTSEG::LoadModelFromFolder(std::string licenseKey, ModelConfig modelConfig, std::string modelName, std::string className, const std::string& modelFolder, std::string& labelMap) {
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
try
|
|
{
|
|
bool result = ANSODBase::LoadModelFromFolder(licenseKey, modelConfig, modelName, className, modelFolder, labelMap);
|
|
if (!result) return false;
|
|
_modelConfig = modelConfig;
|
|
_modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
|
|
_modelConfig.modelType = ModelType::RTSEG;
|
|
_modelConfig.inpHeight = 640;
|
|
_modelConfig.precisionType = PrecisionType::FP32; // Default to FP16 for TensorRT models
|
|
if (_modelConfig.numKPS <= 0 || _modelConfig.numKPS > 133) // 133 = COCO wholebody max
|
|
_modelConfig.numKPS = 17;
|
|
_modelConfig.inpWidth = 640;
|
|
if (_modelConfig.modelMNSThreshold < 0.2)
|
|
_modelConfig.modelMNSThreshold = 0.5;
|
|
if (_modelConfig.modelConfThreshold < 0.2)
|
|
_modelConfig.modelConfThreshold = 0.5;
|
|
if (_modelConfig.kpsThreshold <= 0)_modelConfig.kpsThreshold = 0.5; // If not define
|
|
_fp16 = true; // Load Model from Here
|
|
// Load Model from Here
|
|
TOP_K = 100;
|
|
SEG_CHANNELS = 32;
|
|
PROBABILITY_THRESHOLD = _modelConfig.detectionScoreThreshold;
|
|
NMS_THRESHOLD = _modelConfig.modelMNSThreshold;
|
|
SEGMENTATION_THRESHOLD = 0.5f;
|
|
SEG_H = 160;
|
|
SEG_W = 160;
|
|
NUM_KPS = _modelConfig.numKPS;
|
|
KPS_THRESHOLD = _modelConfig.kpsThreshold;
|
|
SEG_CHANNELS = 32; // For segmentation
|
|
std::string _modelName = modelName;
|
|
if (_modelName.empty()) {
|
|
_modelName = "train_last";
|
|
}
|
|
std::string modelFullName = _modelName + ".onnx";
|
|
if (!m_trtEngine) {
|
|
// Fixed batch size of 1 for this model
|
|
m_options.optBatchSize = _modelConfig.gpuOptBatchSize;
|
|
m_options.maxBatchSize = _modelConfig.gpuMaxBatchSize;
|
|
m_options.deviceIndex = _modelConfig.gpuDeviceIndex;
|
|
m_options.maxInputHeight = _modelConfig.maxInputHeight;
|
|
m_options.minInputHeight = _modelConfig.minInputHeight;
|
|
m_options.optInputHeight = _modelConfig.optInputHeight;
|
|
m_options.maxInputWidth = _modelConfig.maxInputWidth;
|
|
m_options.minInputWidth = _modelConfig.minInputWidth;
|
|
m_options.optInputWidth = _modelConfig.optInputWidth;
|
|
m_options.engineFileDir = _modelFolder;
|
|
// Use FP16 or FP32 precision based on the input flag
|
|
m_options.precision = (_fp16 ? Precision::FP16 : Precision::FP32);
|
|
// Create the TensorRT inference engine
|
|
m_trtEngine = std::make_unique<Engine<float>>(m_options);
|
|
}
|
|
// 0. Check if the configuration file exist
|
|
if (FileExist(_modelConfigFile)) {
|
|
ModelType modelType;
|
|
std::vector<int> inputShape;
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
|
if (inputShape.size() == 2) {
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
|
}
|
|
}
|
|
else {// This is old version of model zip file
|
|
_modelFilePath = CreateFilePath(_modelFolder, modelFullName);
|
|
_classFilePath = CreateFilePath(_modelFolder, className);
|
|
std::ifstream isValidFileName(_classFilePath);
|
|
if (!isValidFileName)
|
|
{
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromString();
|
|
}
|
|
else {
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromFile();
|
|
}
|
|
}
|
|
// 1. Load labelMap and engine
|
|
labelMap.clear();
|
|
if (!_classes.empty())
|
|
labelMap = VectorToCommaSeparatedString(_classes);
|
|
|
|
// Load the TensorRT engine file
|
|
if (this->_loadEngineOnCreation) {
|
|
auto succ = m_trtEngine->buildLoadNetwork(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE, m_maxSlotsPerGpu);
|
|
if (!succ) {
|
|
const std::string errMsg = "Error: Unable to load TensorRT engine weights into memory. " + _modelFilePath;
|
|
this->_logger.LogError("TENSORRTSEG::Initialize", errMsg, __FILE__, __LINE__);
|
|
_modelLoadValid = false;
|
|
return false;
|
|
}
|
|
}
|
|
_modelLoadValid = true;
|
|
_isInitialized = true;
|
|
return true;
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::LoadModelFromFolder", e.what(), __FILE__, __LINE__);
|
|
return false;
|
|
}
|
|
}
|
|
bool TENSORRTSEG::Initialize(std::string licenseKey, ModelConfig modelConfig, const std::string& modelZipFilePath, const std::string& modelZipPassword, std::string& labelMap) {
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
try {
|
|
const bool engineAlreadyLoaded = _modelLoadValid && _isInitialized && m_trtEngine != nullptr;
|
|
_modelLoadValid = false;
|
|
bool result = ANSODBase::Initialize(licenseKey, modelConfig, modelZipFilePath, modelZipPassword, labelMap);
|
|
if (!result) return false;
|
|
// Parsing for YOLO only here
|
|
_modelConfig = modelConfig;
|
|
_modelConfig.detectionType = ANSCENTER::DetectionType::SEGMENTATION;
|
|
_modelConfig.modelType = ModelType::RTSEG;;
|
|
_modelConfig.inpHeight = 640;
|
|
_modelConfig.inpWidth = 640;
|
|
if (_modelConfig.numKPS <= 0 || _modelConfig.numKPS > 133) // 133 = COCO wholebody max
|
|
_modelConfig.numKPS = 17;
|
|
_modelConfig.precisionType = PrecisionType::FP32; // Default to FP16 for TensorRT models
|
|
if (_modelConfig.modelMNSThreshold < 0.2)
|
|
_modelConfig.modelMNSThreshold = 0.5;
|
|
if (_modelConfig.modelConfThreshold < 0.2)
|
|
_modelConfig.modelConfThreshold = 0.5;
|
|
if (_modelConfig.kpsThreshold == 0)_modelConfig.kpsThreshold = 0.5; // If not define
|
|
_fp16 = true; // Load Model from Here
|
|
// Load Model from Here
|
|
TOP_K = 100;
|
|
SEG_CHANNELS = 32;
|
|
PROBABILITY_THRESHOLD = _modelConfig.detectionScoreThreshold;
|
|
NMS_THRESHOLD = _modelConfig.modelMNSThreshold;
|
|
SEGMENTATION_THRESHOLD = 0.5f;
|
|
SEG_H = 160;
|
|
SEG_W = 160;
|
|
NUM_KPS = _modelConfig.numKPS;
|
|
KPS_THRESHOLD = _modelConfig.kpsThreshold;
|
|
SEG_CHANNELS = 32; // For segmentation
|
|
|
|
if (!m_trtEngine) {
|
|
// Fixed batch size of 1 for this model
|
|
m_options.optBatchSize = _modelConfig.gpuOptBatchSize;
|
|
m_options.maxBatchSize = _modelConfig.gpuMaxBatchSize;
|
|
m_options.deviceIndex = _modelConfig.gpuDeviceIndex;
|
|
m_options.maxInputHeight = _modelConfig.maxInputHeight;
|
|
m_options.minInputHeight = _modelConfig.minInputHeight;
|
|
m_options.optInputHeight = _modelConfig.optInputHeight;
|
|
m_options.maxInputWidth = _modelConfig.maxInputWidth;
|
|
m_options.minInputWidth = _modelConfig.minInputWidth;
|
|
m_options.optInputWidth = _modelConfig.optInputWidth;
|
|
m_options.engineFileDir = _modelFolder;
|
|
// Use FP16 or FP32 precision based on the input flag
|
|
m_options.precision = (_fp16 ? Precision::FP16 : Precision::FP32);
|
|
// Create the TensorRT inference engine
|
|
m_trtEngine = std::make_unique<Engine<float>>(m_options);
|
|
}
|
|
// 0. Check if the configuration file exist
|
|
if (FileExist(_modelConfigFile)) {
|
|
ModelType modelType;
|
|
std::vector<int> inputShape;
|
|
_classes = ANSUtilityHelper::GetConfigFileContent(_modelConfigFile, modelType, inputShape);
|
|
if (inputShape.size() == 2) {
|
|
if (inputShape[0] > 0)_modelConfig.inpHeight = inputShape[0];
|
|
if (inputShape[1] > 0)_modelConfig.inpWidth = inputShape[1];
|
|
}
|
|
}
|
|
else {// This is old version of model zip file
|
|
_modelFilePath = CreateFilePath(_modelFolder, "train_last.onnx");
|
|
_classFilePath = CreateFilePath(_modelFolder, "classes.names");
|
|
std::ifstream isValidFileName(_classFilePath);
|
|
if (!isValidFileName)
|
|
{
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from string", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromString();
|
|
}
|
|
else {
|
|
this->_logger.LogDebug("TENSORRTSEG::Initialize. Load classes from file", _classFilePath, __FILE__, __LINE__);
|
|
LoadClassesFromFile();
|
|
}
|
|
}
|
|
// 1. Load labelMap and engine
|
|
labelMap.clear();
|
|
if (!_classes.empty())
|
|
labelMap = VectorToCommaSeparatedString(_classes);
|
|
|
|
// Load the TensorRT engine file
|
|
if (this->_loadEngineOnCreation && !engineAlreadyLoaded) {
|
|
auto succ = m_trtEngine->buildLoadNetwork(_modelFilePath, SUB_VALS, DIV_VALS, NORMALIZE, m_maxSlotsPerGpu);
|
|
if (!succ) {
|
|
const std::string errMsg = "Error: Unable to load TensorRT engine weights into memory. " + _modelFilePath;
|
|
this->_logger.LogError("TENSORRTSEG::Initialize", errMsg, __FILE__, __LINE__);
|
|
_modelLoadValid = false;
|
|
return false;
|
|
}
|
|
}
|
|
_modelLoadValid = true;
|
|
_isInitialized = true;
|
|
return true;
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::Initialize", e.what(), __FILE__, __LINE__);
|
|
return false;
|
|
}
|
|
}
|
|
std::vector<Object> TENSORRTSEG::RunInference(const cv::Mat& inputImgBGR) {
|
|
return RunInference(inputImgBGR, "");
|
|
}
|
|
std::vector<Object> TENSORRTSEG::RunInference(const cv::Mat& inputImgBGR,const std::string& camera_id)
|
|
{
|
|
{
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
if (!_modelLoadValid) {
|
|
_logger.LogError("TENSORRTSEG::RunInference",
|
|
"Cannot load TensorRT model", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (!_licenseValid) {
|
|
_logger.LogError("TENSORRTSEG::RunInference",
|
|
"Invalid license", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (!_isInitialized) {
|
|
_logger.LogError("TENSORRTSEG::RunInference",
|
|
"Model not initialized", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (inputImgBGR.empty() || inputImgBGR.cols < 10 || inputImgBGR.rows < 10) {
|
|
return {};
|
|
}
|
|
}
|
|
try {
|
|
return DetectObjects(inputImgBGR, camera_id);
|
|
}
|
|
catch (const std::exception& e) {
|
|
_logger.LogFatal("TENSORRTSEG::RunInference", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
|
|
TENSORRTSEG::~TENSORRTSEG() {
|
|
try {
|
|
Destroy();
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogError("TENSORRTSEG::~TENSORRTSEG()", e.what(), __FILE__, __LINE__);
|
|
}
|
|
}
|
|
bool TENSORRTSEG::Destroy() {
|
|
try {
|
|
m_trtEngine.reset(); // Releases the current engine and sets m_trtEngine to nullptr.
|
|
return true;
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogError("TENSORRTSEG::~TENSORRTSEG()", e.what(), __FILE__, __LINE__);
|
|
return false;
|
|
}
|
|
}
|
|
// private
|
|
std::vector<cv::Point2f> TENSORRTSEG::maskToPolygon(const cv::Mat& binaryMask,const cv::Rect& boundingBox,float simplificationEpsilon,int minContourArea)
|
|
{
|
|
std::vector<cv::Point2f> polygon;
|
|
|
|
try {
|
|
// Validate input
|
|
if (binaryMask.empty() || binaryMask.type() != CV_8UC1) {
|
|
return polygon;
|
|
}
|
|
|
|
// Extract region of interest from mask
|
|
cv::Rect roi = boundingBox & cv::Rect(0, 0, binaryMask.cols, binaryMask.rows);
|
|
if (roi.area() <= 0) {
|
|
return polygon;
|
|
}
|
|
|
|
cv::Mat maskROI = binaryMask(roi);
|
|
|
|
// Find contours in the mask
|
|
std::vector<std::vector<cv::Point>> contours;
|
|
std::vector<cv::Vec4i> hierarchy;
|
|
cv::findContours(maskROI.clone(), contours, hierarchy,
|
|
cv::RETR_EXTERNAL, cv::CHAIN_APPROX_SIMPLE);
|
|
|
|
if (contours.empty()) {
|
|
return polygon;
|
|
}
|
|
|
|
// Find the largest contour (main object)
|
|
int largestIdx = 0;
|
|
double largestArea = 0.0;
|
|
|
|
for (size_t i = 0; i < contours.size(); ++i) {
|
|
double area = cv::contourArea(contours[i]);
|
|
if (area > largestArea && area >= minContourArea) {
|
|
largestArea = area;
|
|
largestIdx = static_cast<int>(i);
|
|
}
|
|
}
|
|
|
|
if (largestArea < minContourArea) {
|
|
return polygon;
|
|
}
|
|
|
|
// Simplify the contour to reduce number of points
|
|
std::vector<cv::Point> simplifiedContour;
|
|
cv::approxPolyDP(contours[largestIdx], simplifiedContour,
|
|
simplificationEpsilon, true);
|
|
|
|
// Convert to Point2f and offset by ROI position
|
|
polygon.reserve(simplifiedContour.size());
|
|
for (const auto& pt : simplifiedContour) {
|
|
polygon.emplace_back(
|
|
static_cast<float>(pt.x + roi.x),
|
|
static_cast<float>(pt.y + roi.y)
|
|
);
|
|
}
|
|
|
|
return polygon;
|
|
}
|
|
catch (const cv::Exception& e) {
|
|
// Log error if logger available
|
|
polygon.clear();
|
|
return polygon;
|
|
}
|
|
}
|
|
std::vector<Object> TENSORRTSEG::DetectObjects(const cv::Mat& inputImage, const std::string& camera_id) {
|
|
try {
|
|
// --- 1. Set GPU device context ---
|
|
if (m_trtEngine) {
|
|
m_trtEngine->setDeviceContext();
|
|
}
|
|
|
|
// --- 1b. CUDA context health check ---
|
|
if (!m_nv12Helper.isCudaContextHealthy(_logger, "TENSORRTSEG")) {
|
|
return {};
|
|
}
|
|
|
|
// --- 2. Preprocess under lock ---
|
|
ImageMetadata meta;
|
|
std::vector<std::vector<cv::cuda::GpuMat>> input;
|
|
bool usedNV12 = false;
|
|
float bgrFullResScaleX = 1.0f, bgrFullResScaleY = 1.0f;
|
|
{
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
const int inferenceGpu = m_trtEngine ? m_trtEngine->getPreferredDeviceIndex() : 0;
|
|
const auto& inputDims = m_trtEngine->getInputDims();
|
|
const int inputW = inputDims[0].d[2];
|
|
const int inputH = inputDims[0].d[1];
|
|
|
|
auto nv12 = m_nv12Helper.tryNV12(inputImage, inferenceGpu, inputW, inputH,
|
|
NV12PreprocessHelper::defaultYOLOLauncher(),
|
|
_logger, "TENSORRTSEG");
|
|
if (nv12.succeeded) {
|
|
meta.imgWidth = nv12.metaWidth;
|
|
meta.imgHeight = nv12.metaHeight;
|
|
meta.ratio = nv12.ratio;
|
|
input = {{ std::move(nv12.gpuRGB) }};
|
|
usedNV12 = true;
|
|
}
|
|
else if (nv12.useBgrFullRes) {
|
|
input = Preprocess(nv12.bgrFullResImg, meta);
|
|
usedNV12 = !input.empty();
|
|
bgrFullResScaleX = nv12.bgrFullResScaleX;
|
|
bgrFullResScaleY = nv12.bgrFullResScaleY;
|
|
}
|
|
|
|
if (input.empty()) {
|
|
input = Preprocess(inputImage, meta);
|
|
}
|
|
m_nv12Helper.tickInference();
|
|
}
|
|
|
|
if (input.empty()) return {};
|
|
|
|
// Phase 2: Inference — mutex released; pool dispatches to idle GPU slot
|
|
std::vector<std::vector<std::vector<float>>> featureVectors;
|
|
auto succ = m_trtEngine->runInference(input, featureVectors);
|
|
if (!succ) {
|
|
this->_logger.LogError("TENSORRTSEG::DetectObjects", "Error running inference", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Phase 3: Postprocess under lock
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
std::vector<std::vector<float>> featureVector;
|
|
Engine<float>::transformOutput(featureVectors, featureVector);
|
|
auto ret = PostProcessSegmentation(featureVector, camera_id, meta);
|
|
|
|
// --- 4b. Rescale coords from full-res to display-res ---
|
|
if (bgrFullResScaleX != 1.0f || bgrFullResScaleY != 1.0f) {
|
|
for (auto& obj : ret) {
|
|
obj.box.x = static_cast<int>(obj.box.x * bgrFullResScaleX);
|
|
obj.box.y = static_cast<int>(obj.box.y * bgrFullResScaleY);
|
|
obj.box.width = static_cast<int>(obj.box.width * bgrFullResScaleX);
|
|
obj.box.height = static_cast<int>(obj.box.height * bgrFullResScaleY);
|
|
for (auto& pt : obj.polygon) {
|
|
pt.x *= bgrFullResScaleX;
|
|
pt.y *= bgrFullResScaleY;
|
|
}
|
|
}
|
|
}
|
|
|
|
if (_trackerEnabled) {
|
|
ret = ApplyTracking(ret, camera_id);
|
|
if (_stabilizationEnabled) ret = StabilizeDetections(ret, camera_id);
|
|
}
|
|
return ret;
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::DetectObjects", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
std::vector<std::vector<cv::cuda::GpuMat>> TENSORRTSEG::Preprocess(const cv::Mat& inputImage, ImageMetadata& outMeta) {
|
|
try {
|
|
if (!_licenseValid) {
|
|
this->_logger.LogFatal("TENSORRTSEG::Preprocess", "Invalid license", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Get model input dimensions
|
|
const auto& inputDims = m_trtEngine->getInputDims();
|
|
const int inputH = inputDims[0].d[1];
|
|
const int inputW = inputDims[0].d[2];
|
|
// Upload the image to GPU memory
|
|
cv::cuda::Stream stream;
|
|
cv::cuda::GpuMat img;
|
|
|
|
if (inputImage.channels() == 1) {
|
|
// Convert grayscale to 3-channel BGR before uploading
|
|
cv::Mat img3Channel;
|
|
cv::cvtColor(inputImage, img3Channel, cv::COLOR_GRAY2BGR);
|
|
img.upload(img3Channel, stream);
|
|
}
|
|
else {
|
|
img.upload(inputImage, stream);
|
|
}
|
|
|
|
// Convert to RGB
|
|
cv::cuda::GpuMat imgRGB;
|
|
cv::cuda::cvtColor(img, imgRGB, cv::COLOR_BGR2RGB, 0, stream);
|
|
stream.waitForCompletion();
|
|
|
|
// Set image size parameters
|
|
outMeta.imgHeight = imgRGB.rows;
|
|
outMeta.imgWidth = imgRGB.cols;
|
|
if (outMeta.imgHeight > 0 && outMeta.imgWidth > 0) {
|
|
outMeta.ratio = 1.f / std::min(inputDims[0].d[2] / static_cast<float>(imgRGB.cols),
|
|
inputDims[0].d[1] / static_cast<float>(imgRGB.rows));
|
|
|
|
cv::cuda::GpuMat resized = imgRGB;
|
|
|
|
// Resize to the model's expected input size while maintaining aspect ratio with padding
|
|
if (resized.rows != inputDims[0].d[1] || resized.cols != inputDims[0].d[2]) {
|
|
resized = Engine<float>::resizeKeepAspectRatioPadRightBottom(imgRGB, inputDims[0].d[1], inputDims[0].d[2]);
|
|
}
|
|
|
|
// Convert to format expected by our inference engine
|
|
std::vector<cv::cuda::GpuMat> input{ std::move(resized) };
|
|
std::vector<std::vector<cv::cuda::GpuMat>> inputs{ std::move(input) };
|
|
return inputs;
|
|
}
|
|
else {
|
|
this->_logger.LogFatal("TENSORRTCL::Preprocess",
|
|
"Image height or width is zero after processing (Width: " + std::to_string(outMeta.imgWidth) +
|
|
", Height: " + std::to_string(outMeta.imgHeight) + ")",
|
|
__FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
catch (const std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::Preprocess", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
std::vector<Object> TENSORRTSEG::PostProcessSegmentation(std::vector<std::vector<float>>& featureVectors, const std::string& camera_id, const ImageMetadata& meta) {
|
|
try {
|
|
if (!_licenseValid) {
|
|
this->_logger.LogFatal("TENSORRTSEG::PostProcessSegmentation", "Invalid license", __FILE__, __LINE__);
|
|
std::vector<Object> result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
const auto& outputDims = m_trtEngine->getOutputDims();
|
|
|
|
int numChannels = outputDims[0].d[1];
|
|
int numAnchors = outputDims[0].d[2];
|
|
|
|
const auto numClasses = numChannels - SEG_CHANNELS - 4;
|
|
|
|
// Ensure the output lengths are correct
|
|
if (featureVectors[0].size() == static_cast<size_t>(numChannels) * numAnchors &&
|
|
featureVectors[1].size() == static_cast<size_t>(SEG_CHANNELS) * SEG_H * SEG_W) {
|
|
|
|
cv::Mat output = cv::Mat(numChannels, numAnchors, CV_32F, featureVectors[0].data());
|
|
output = output.t();
|
|
|
|
cv::Mat protos = cv::Mat(SEG_CHANNELS, SEG_H * SEG_W, CV_32F, featureVectors[1].data());
|
|
|
|
std::vector<int> labels;
|
|
std::vector<float> scores;
|
|
std::vector<cv::Rect> bboxes;
|
|
std::vector<cv::Mat> maskConfs;
|
|
std::vector<int> indices;
|
|
|
|
// Object the bounding boxes and class labels
|
|
for (int i = 0; i < numAnchors; i++) {
|
|
auto rowPtr = output.row(i).ptr<float>();
|
|
auto bboxesPtr = rowPtr;
|
|
auto scoresPtr = rowPtr + 4;
|
|
auto maskConfsPtr = rowPtr + 4 + numClasses;
|
|
auto maxSPtr = std::max_element(scoresPtr, scoresPtr + numClasses);
|
|
float score = *maxSPtr;
|
|
if (score > this->_modelConfig.detectionScoreThreshold)
|
|
{
|
|
float x = *bboxesPtr++;
|
|
float y = *bboxesPtr++;
|
|
float w = *bboxesPtr++;
|
|
float h = *bboxesPtr;
|
|
|
|
float x0 = std::clamp((x - 0.5f * w) * meta.ratio, 0.f, meta.imgWidth);
|
|
float y0 = std::clamp((y - 0.5f * h) * meta.ratio, 0.f, meta.imgHeight);
|
|
float x1 = std::clamp((x + 0.5f * w) * meta.ratio, 0.f, meta.imgWidth);
|
|
float y1 = std::clamp((y + 0.5f * h) * meta.ratio, 0.f, meta.imgHeight);
|
|
|
|
int label = maxSPtr - scoresPtr;
|
|
cv::Rect_<float> bbox;
|
|
bbox.x = x0;
|
|
bbox.y = y0;
|
|
bbox.width = x1 - x0;
|
|
bbox.height = y1 - y0;
|
|
bbox.x = std::max(0.f, bbox.x);
|
|
bbox.y = std::max(0.f, bbox.y);
|
|
bbox.width = std::min(meta.imgWidth - bbox.x, bbox.width);
|
|
bbox.height = std::min(meta.imgHeight - bbox.y, bbox.height);
|
|
cv::Mat maskConf = cv::Mat(1, SEG_CHANNELS, CV_32F, maskConfsPtr);
|
|
bboxes.push_back(bbox);
|
|
labels.push_back(label);
|
|
scores.push_back(score);
|
|
maskConfs.push_back(maskConf);
|
|
}
|
|
}
|
|
cv::dnn::NMSBoxesBatched(bboxes, scores, labels, PROBABILITY_THRESHOLD, NMS_THRESHOLD, indices);
|
|
cv::Mat masks;
|
|
int classNameSize = static_cast<int>(_classes.size());
|
|
std::vector<Object> objs;
|
|
for (auto& i : indices) {
|
|
if (scores[i] > _modelConfig.detectionScoreThreshold) {
|
|
cv::Rect tmp = bboxes[i];
|
|
Object obj;
|
|
obj.classId = labels[i];
|
|
if (!_classes.empty()) {
|
|
if (obj.classId < classNameSize) {
|
|
obj.className = _classes[obj.classId];
|
|
}
|
|
else {
|
|
obj.className = _classes[classNameSize - 1]; // Use last valid class name if out of range
|
|
}
|
|
}
|
|
else {
|
|
obj.className = "Unknown"; // Fallback if _classes is empty
|
|
}
|
|
obj.box = tmp;
|
|
obj.confidence = scores[i];
|
|
obj.className = _classes[labels[i]];
|
|
masks.push_back(maskConfs[i]);
|
|
objs.push_back(obj);
|
|
}
|
|
}
|
|
if (!masks.empty()) {
|
|
cv::Mat matmulRes = (masks * protos).t();
|
|
cv::Mat maskMat = matmulRes.reshape(indices.size(), { _modelConfig.inpWidth, _modelConfig.inpHeight });
|
|
std::vector<cv::Mat> maskChannels;
|
|
cv::split(maskMat, maskChannels);
|
|
const auto inputDims = m_trtEngine->getInputDims();
|
|
|
|
cv::Rect roi;
|
|
if (meta.imgHeight > meta.imgWidth) {
|
|
roi = cv::Rect(0, 0, _modelConfig.inpWidth * meta.imgWidth / meta.imgHeight, _modelConfig.inpHeight);
|
|
}
|
|
else {
|
|
roi = cv::Rect(0, 0, _modelConfig.inpWidth, _modelConfig.inpHeight * meta.imgHeight / meta.imgWidth);
|
|
}
|
|
for (size_t i = 0; i < indices.size(); i++)
|
|
{
|
|
cv::Mat dest, mask;
|
|
cv::exp(-maskChannels[i], dest);
|
|
dest = 1.0 / (1.0 + dest);
|
|
dest = dest(roi);
|
|
objs[i].cameraId = camera_id;
|
|
cv::resize(
|
|
dest,
|
|
mask,
|
|
cv::Size(static_cast<int>(meta.imgWidth), static_cast<int>(meta.imgHeight)),
|
|
cv::INTER_LINEAR
|
|
);
|
|
objs[i].mask = mask(objs[i].box) > _modelConfig.modelConfThreshold;// Need to check segmentation
|
|
objs[i].polygon = maskToPolygon(objs[i].mask, objs[i].box, 2.0f, 10);
|
|
|
|
// Alternative: Get multiple polygons if needed
|
|
// auto polygons = maskToPolygons(finalMask, result.box, 2.0f, 10, 3);
|
|
// result.polygon = polygons.empty() ? std::vector<cv::Point2f>() : polygons[0];
|
|
|
|
// Validate polygon
|
|
if (objs[i].polygon.size() < 3) {
|
|
// Fallback to bounding box if polygon extraction failed
|
|
objs[i].polygon = {
|
|
cv::Point2f(objs[i].box.x, objs[i].box.y),
|
|
cv::Point2f(objs[i].box.x + objs[i].box.width, objs[i].box.y),
|
|
cv::Point2f(objs[i].box.x + objs[i].box.width, objs[i].box.y + objs[i].box.height),
|
|
cv::Point2f(objs[i].box.x, objs[i].box.y + objs[i].box.height)
|
|
};
|
|
}
|
|
}
|
|
}
|
|
//EnqueueDetection(objs, camera_id);
|
|
return objs;
|
|
}
|
|
else {
|
|
return std::vector<Object>();
|
|
}
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::PostProcessSegmentation", e.what(), __FILE__, __LINE__);
|
|
std::vector<Object>result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
}
|
|
|
|
std::vector<std::vector<Object>> TENSORRTSEG::DetectObjectsBatch(const std::vector<cv::Mat>& inputImages, const std::string& camera_id) {
|
|
{
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
if (inputImages.empty()) {
|
|
_logger.LogFatal("TENSORRTSEG::DetectObjectsBatch", "Empty input images vector", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
|
|
// Auto-split if batch exceeds engine capacity
|
|
const int maxBatch = m_options.maxBatchSize > 0 ? m_options.maxBatchSize : 1;
|
|
if (static_cast<int>(inputImages.size()) > maxBatch) {
|
|
const size_t numImages = inputImages.size();
|
|
std::vector<std::vector<Object>> allResults;
|
|
allResults.reserve(numImages);
|
|
// Process chunks sequentially to avoid GPU contention on the same engine
|
|
for (size_t start = 0; start < numImages; start += static_cast<size_t>(maxBatch)) {
|
|
const size_t end = std::min(start + static_cast<size_t>(maxBatch), numImages);
|
|
std::vector<cv::Mat> chunk(inputImages.begin() + start, inputImages.begin() + end);
|
|
auto chunkResults = DetectObjectsBatch(chunk, camera_id);
|
|
if (chunkResults.size() == chunk.size()) {
|
|
for (auto& r : chunkResults) allResults.push_back(std::move(r));
|
|
}
|
|
else {
|
|
_logger.LogError("TENSORRTSEG::DetectObjectsBatch",
|
|
"Chunk returned " + std::to_string(chunkResults.size()) +
|
|
" results, expected " + std::to_string(chunk.size()) +
|
|
". Padding with empty results.", __FILE__, __LINE__);
|
|
for (auto& r : chunkResults) allResults.push_back(std::move(r));
|
|
for (size_t pad = chunkResults.size(); pad < chunk.size(); ++pad) {
|
|
allResults.push_back({});
|
|
}
|
|
}
|
|
}
|
|
return allResults;
|
|
}
|
|
|
|
_logger.LogDebug("TENSORRTSEG::DetectObjectsBatch",
|
|
"Processing batch of " + std::to_string(inputImages.size()) + " images",
|
|
__FILE__, __LINE__);
|
|
|
|
// Phase 1: Preprocess under brief lock
|
|
BatchMetadata metadata;
|
|
std::vector<std::vector<cv::cuda::GpuMat>> inputs;
|
|
{
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
inputs = PreprocessBatch(inputImages, metadata);
|
|
}
|
|
if (inputs.empty() || inputs[0].empty()) {
|
|
_logger.LogFatal("TENSORRTSEG::DetectObjectsBatch", "Preprocessing failed", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Phase 2: Inference -- mutex released; pool dispatches to idle GPU slot
|
|
std::vector<std::vector<std::vector<float>>> featureVectors;
|
|
auto succ = m_trtEngine->runInference(inputs, featureVectors);
|
|
if (!succ) {
|
|
_logger.LogError("TENSORRTSEG::DetectObjectsBatch", "Error running inference", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Phase 3: Parallel postprocessing -- each image is independent
|
|
const size_t numBatch = featureVectors.size();
|
|
std::vector<std::vector<Object>> batchDetections(numBatch);
|
|
std::vector<std::future<std::vector<Object>>> postFutures;
|
|
postFutures.reserve(numBatch);
|
|
for (size_t batchIdx = 0; batchIdx < numBatch; ++batchIdx) {
|
|
const auto& batchOutput = featureVectors[batchIdx];
|
|
std::vector<std::vector<float>> fv =
|
|
batchOutput.empty() ? std::vector<std::vector<float>>{} : batchOutput;
|
|
postFutures.push_back(std::async(std::launch::async,
|
|
[this, fv = std::move(fv), cid = camera_id,
|
|
idx = batchIdx, &metadata]() mutable {
|
|
return PostProcessSegmentationBatch(fv, cid, idx, metadata);
|
|
}));
|
|
}
|
|
for (size_t i = 0; i < numBatch; ++i)
|
|
batchDetections[i] = postFutures[i].get();
|
|
|
|
if (_trackerEnabled) {
|
|
for (auto& dets : batchDetections) {
|
|
if (!dets.empty()) {
|
|
dets = ApplyTracking(dets, camera_id);
|
|
if (_stabilizationEnabled) dets = StabilizeDetections(dets, camera_id);
|
|
}
|
|
}
|
|
}
|
|
|
|
_logger.LogDebug("TENSORRTSEG::DetectObjectsBatch",
|
|
"Batch processing complete. Images: " + std::to_string(numBatch),
|
|
__FILE__, __LINE__);
|
|
return batchDetections;
|
|
}
|
|
std::vector<std::vector<cv::cuda::GpuMat>> TENSORRTSEG::PreprocessBatch(const std::vector<cv::Mat>& inputImages, BatchMetadata& outMetadata) {
|
|
try {
|
|
if (!_licenseValid) {
|
|
_logger.LogFatal("TENSORRTSEG::PreprocessBatch", "Invalid license", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
const auto& inputDims = m_trtEngine->getInputDims();
|
|
const int inputH = inputDims[0].d[1];
|
|
const int inputW = inputDims[0].d[2];
|
|
|
|
// Store original image dimensions for each image in batch
|
|
outMetadata.imgHeights.resize(inputImages.size());
|
|
outMetadata.imgWidths.resize(inputImages.size());
|
|
outMetadata.ratios.resize(inputImages.size());
|
|
|
|
std::vector<cv::cuda::GpuMat> batchProcessed;
|
|
batchProcessed.reserve(inputImages.size());
|
|
|
|
cv::cuda::Stream stream;
|
|
|
|
// Process each image
|
|
for (size_t i = 0; i < inputImages.size(); ++i) {
|
|
const auto& inputImage = inputImages[i];
|
|
|
|
if (inputImage.empty()) {
|
|
_logger.LogFatal("TENSORRTSEG::PreprocessBatch",
|
|
"Empty input image at index " + std::to_string(i), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Upload to GPU
|
|
cv::cuda::GpuMat img;
|
|
|
|
// Convert grayscale to BGR if needed
|
|
if (inputImage.channels() == 1) {
|
|
cv::Mat img3Channel;
|
|
cv::cvtColor(inputImage, img3Channel, cv::COLOR_GRAY2BGR);
|
|
img.upload(img3Channel, stream);
|
|
}
|
|
else {
|
|
img.upload(inputImage, stream);
|
|
}
|
|
|
|
// Convert to RGB
|
|
cv::cuda::GpuMat imgRGB;
|
|
cv::cuda::cvtColor(img, imgRGB, cv::COLOR_BGR2RGB, 0, stream);
|
|
|
|
// Store original dimensions
|
|
outMetadata.imgHeights[i] = imgRGB.rows;
|
|
outMetadata.imgWidths[i] = imgRGB.cols;
|
|
|
|
if (outMetadata.imgHeights[i] <= 0 || outMetadata.imgWidths[i] <= 0) {
|
|
_logger.LogFatal("TENSORRTSEG::PreprocessBatch",
|
|
"Image " + std::to_string(i) + " has invalid dimensions (Width: " +
|
|
std::to_string(outMetadata.imgWidths[i]) + ", Height: " +
|
|
std::to_string(outMetadata.imgHeights[i]) + ")",
|
|
__FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
// Calculate ratio for this image
|
|
outMetadata.ratios[i] = 1.f / std::min(inputW / static_cast<float>(imgRGB.cols),
|
|
inputH / static_cast<float>(imgRGB.rows));
|
|
|
|
// Resize with padding
|
|
cv::cuda::GpuMat resized = imgRGB;
|
|
if (resized.rows != inputH || resized.cols != inputW) {
|
|
resized = Engine<float>::resizeKeepAspectRatioPadRightBottom(imgRGB, inputH, inputW);
|
|
}
|
|
|
|
batchProcessed.push_back(std::move(resized));
|
|
}
|
|
|
|
stream.waitForCompletion();
|
|
|
|
// Return as required format
|
|
std::vector<std::vector<cv::cuda::GpuMat>> inputs;
|
|
inputs.push_back(std::move(batchProcessed));
|
|
|
|
return inputs;
|
|
}
|
|
catch (const std::exception& e) {
|
|
_logger.LogFatal("TENSORRTSEG::PreprocessBatch", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
std::vector<Object> TENSORRTSEG::PostProcessSegmentationBatch(std::vector<std::vector<float>>& featureVectors, const std::string& camera_id, size_t batchIdx, const BatchMetadata& metadata) {
|
|
try {
|
|
if (!_licenseValid) {
|
|
_logger.LogFatal("TENSORRTSEG::PostProcessSegmentationBatch", "Invalid license", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
const auto& outputDims = m_trtEngine->getOutputDims();
|
|
|
|
int numChannels = outputDims[0].d[1];
|
|
int numAnchors = outputDims[0].d[2];
|
|
|
|
const auto numClasses = numChannels - SEG_CHANNELS - 4;
|
|
|
|
// Get batch-specific dimensions
|
|
float ratio = metadata.ratios[batchIdx];
|
|
float imgWidth = static_cast<float>(metadata.imgWidths[batchIdx]);
|
|
float imgHeight = static_cast<float>(metadata.imgHeights[batchIdx]);
|
|
|
|
// Ensure the output lengths are correct
|
|
if (featureVectors.size() < 2) {
|
|
_logger.LogError("TENSORRTSEG::PostProcessSegmentationBatch",
|
|
"Invalid feature vectors size: " + std::to_string(featureVectors.size()),
|
|
__FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
if (featureVectors[0].size() != static_cast<size_t>(numChannels) * numAnchors) {
|
|
_logger.LogError("TENSORRTSEG::PostProcessSegmentationBatch",
|
|
"Detection output size mismatch", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
if (featureVectors[1].size() != static_cast<size_t>(SEG_CHANNELS) * SEG_H * SEG_W) {
|
|
_logger.LogError("TENSORRTSEG::PostProcessSegmentationBatch",
|
|
"Segmentation mask size mismatch", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
|
|
cv::Mat output = cv::Mat(numChannels, numAnchors, CV_32F, featureVectors[0].data());
|
|
output = output.t();
|
|
|
|
cv::Mat protos = cv::Mat(SEG_CHANNELS, SEG_H * SEG_W, CV_32F, featureVectors[1].data());
|
|
|
|
std::vector<int> labels;
|
|
std::vector<float> scores;
|
|
std::vector<cv::Rect> bboxes;
|
|
std::vector<cv::Mat> maskConfs;
|
|
std::vector<int> indices;
|
|
|
|
// Extract bounding boxes and class labels
|
|
for (int i = 0; i < numAnchors; i++) {
|
|
auto rowPtr = output.row(i).ptr<float>();
|
|
auto bboxesPtr = rowPtr;
|
|
auto scoresPtr = rowPtr + 4;
|
|
auto maskConfsPtr = rowPtr + 4 + numClasses;
|
|
auto maxSPtr = std::max_element(scoresPtr, scoresPtr + numClasses);
|
|
float score = *maxSPtr;
|
|
|
|
if (score > _modelConfig.detectionScoreThreshold) {
|
|
float x = *bboxesPtr++;
|
|
float y = *bboxesPtr++;
|
|
float w = *bboxesPtr++;
|
|
float h = *bboxesPtr;
|
|
|
|
// Use batch-specific ratio and dimensions
|
|
float x0 = std::clamp((x - 0.5f * w) * ratio, 0.f, imgWidth);
|
|
float y0 = std::clamp((y - 0.5f * h) * ratio, 0.f, imgHeight);
|
|
float x1 = std::clamp((x + 0.5f * w) * ratio, 0.f, imgWidth);
|
|
float y1 = std::clamp((y + 0.5f * h) * ratio, 0.f, imgHeight);
|
|
|
|
int label = maxSPtr - scoresPtr;
|
|
cv::Rect_<float> bbox;
|
|
bbox.x = x0;
|
|
bbox.y = y0;
|
|
bbox.width = x1 - x0;
|
|
bbox.height = y1 - y0;
|
|
|
|
// Clamp to image boundaries
|
|
bbox.x = std::max(0.f, bbox.x);
|
|
bbox.y = std::max(0.f, bbox.y);
|
|
bbox.width = std::min(imgWidth - bbox.x, bbox.width);
|
|
bbox.height = std::min(imgHeight - bbox.y, bbox.height);
|
|
|
|
cv::Mat maskConf = cv::Mat(1, SEG_CHANNELS, CV_32F, maskConfsPtr);
|
|
bboxes.push_back(bbox);
|
|
labels.push_back(label);
|
|
scores.push_back(score);
|
|
maskConfs.push_back(maskConf);
|
|
}
|
|
}
|
|
|
|
// Run NMS
|
|
cv::dnn::NMSBoxesBatched(bboxes, scores, labels, PROBABILITY_THRESHOLD, NMS_THRESHOLD, indices);
|
|
|
|
cv::Mat masks;
|
|
int classNameSize = static_cast<int>(_classes.size());
|
|
std::vector<Object> objs;
|
|
|
|
for (auto& i : indices) {
|
|
if (scores[i] > _modelConfig.detectionScoreThreshold) {
|
|
cv::Rect tmp = bboxes[i];
|
|
Object obj;
|
|
obj.classId = labels[i];
|
|
|
|
if (!_classes.empty()) {
|
|
if (obj.classId < classNameSize) {
|
|
obj.className = _classes[obj.classId];
|
|
}
|
|
else {
|
|
obj.className = _classes[classNameSize - 1];
|
|
}
|
|
}
|
|
else {
|
|
obj.className = "Unknown";
|
|
}
|
|
|
|
obj.box = tmp;
|
|
obj.confidence = scores[i];
|
|
masks.push_back(maskConfs[i]);
|
|
objs.push_back(obj);
|
|
}
|
|
}
|
|
|
|
// Process segmentation masks
|
|
if (!masks.empty()) {
|
|
cv::Mat matmulRes = (masks * protos).t();
|
|
cv::Mat maskMat = matmulRes.reshape(indices.size(), { _modelConfig.inpWidth, _modelConfig.inpHeight });
|
|
std::vector<cv::Mat> maskChannels;
|
|
cv::split(maskMat, maskChannels);
|
|
|
|
cv::Rect roi;
|
|
if (imgHeight > imgWidth) {
|
|
roi = cv::Rect(0, 0, _modelConfig.inpWidth * imgWidth / imgHeight, _modelConfig.inpHeight);
|
|
}
|
|
else {
|
|
roi = cv::Rect(0, 0, _modelConfig.inpWidth, _modelConfig.inpHeight * imgHeight / imgWidth);
|
|
}
|
|
|
|
for (size_t i = 0; i < indices.size(); i++) {
|
|
cv::Mat dest, mask;
|
|
cv::exp(-maskChannels[i], dest);
|
|
dest = 1.0 / (1.0 + dest);
|
|
dest = dest(roi);
|
|
objs[i].cameraId = camera_id;
|
|
|
|
cv::resize(
|
|
dest,
|
|
mask,
|
|
cv::Size(static_cast<int>(imgWidth), static_cast<int>(imgHeight)),
|
|
cv::INTER_LINEAR
|
|
);
|
|
|
|
objs[i].mask = mask(objs[i].box) > _modelConfig.modelConfThreshold;
|
|
objs[i].polygon = maskToPolygon(objs[i].mask, objs[i].box, 2.0f, 10);
|
|
|
|
// Validate polygon
|
|
if (objs[i].polygon.size() < 3) {
|
|
// Fallback to bounding box if polygon extraction failed
|
|
objs[i].polygon = {
|
|
cv::Point2f(objs[i].box.x, objs[i].box.y),
|
|
cv::Point2f(objs[i].box.x + objs[i].box.width, objs[i].box.y),
|
|
cv::Point2f(objs[i].box.x + objs[i].box.width, objs[i].box.y + objs[i].box.height),
|
|
cv::Point2f(objs[i].box.x, objs[i].box.y + objs[i].box.height)
|
|
};
|
|
}
|
|
}
|
|
}
|
|
|
|
return objs;
|
|
}
|
|
catch (std::exception& e) {
|
|
_logger.LogFatal("TENSORRTSEG::PostProcessSegmentationBatch", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
std::vector<std::vector<Object>> TENSORRTSEG::RunInferencesBatch(
|
|
const std::vector<cv::Mat>& inputs, const std::string& camera_id)
|
|
{
|
|
{
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
if (!_modelLoadValid) {
|
|
_logger.LogError("TENSORRTSEG::RunInferencesBatch",
|
|
"Model not loaded", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (!_licenseValid) {
|
|
_logger.LogError("TENSORRTSEG::RunInferencesBatch",
|
|
"Invalid license", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (!_isInitialized) {
|
|
_logger.LogError("TENSORRTSEG::RunInferencesBatch",
|
|
"Engine not initialized", __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
if (inputs.empty()) return {};
|
|
}
|
|
try {
|
|
return DetectObjectsBatch(inputs, camera_id);
|
|
}
|
|
catch (const std::exception& e) {
|
|
_logger.LogFatal("TENSORRTSEG::RunInferencesBatch", e.what(), __FILE__, __LINE__);
|
|
return {};
|
|
}
|
|
}
|
|
}
|
|
|
|
|
|
/*std::vector<Object> TENSORRTSEG::RunInference(const cv::Mat& inputImgBGR, const std::string& camera_id) {
|
|
std::lock_guard<std::recursive_mutex> lock(_mutex);
|
|
if (!_modelLoadValid) {
|
|
this->_logger.LogFatal("TENSORRTSEG::RunInference", "Cannot load the TensorRT model. Please check if it is exist", __FILE__, __LINE__);
|
|
std::vector<Object> result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
if (!_licenseValid) {
|
|
this->_logger.LogFatal("TENSORRTSEG::RunInference", "Runtime license is not valid or expired. Please contact ANSCENTER", __FILE__, __LINE__);
|
|
std::vector<Object> result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
if (!_isInitialized) {
|
|
this->_logger.LogFatal("TENSORRTSEG::RunInference", "Model is not initialized", __FILE__, __LINE__);
|
|
std::vector<Object> result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
try {
|
|
std::vector<Object> result;
|
|
if (inputImgBGR.empty()) return result;
|
|
if ((inputImgBGR.cols < 10) || (inputImgBGR.rows < 10)) return result;
|
|
return DetectObjects(inputImgBGR, camera_id);
|
|
}
|
|
catch (std::exception& e) {
|
|
this->_logger.LogFatal("TENSORRTSEG::RunInference", e.what(), __FILE__, __LINE__);
|
|
std::vector<Object> result;
|
|
result.clear();
|
|
return result;
|
|
}
|
|
}*/ |