mirrors/monado
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/monado/monado.git synced 2025-02-05 13:28:16 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

d/ht: fix stack overflow on hand detection

Browse source
This commit is contained in:
Simon Zeni 2021-09-30 13:56:36 -04:00
parent ad7b5cb786
commit 036eb19343
1 changed files with 24 additions and 22 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

46
src/xrt/drivers/ht/ht_models.hpp
View file

@ -435,30 +435,30 @@ static std::vector<std::vector<float>> anchor{
static Hand2D
runKeypointEstimator(struct ht_view *htv, cv::Mat img)
{
const size_t lix = 224;
const size_t liy = 224;
cv::Mat planes[3];
constexpr size_t lix = 224;
constexpr size_t liy = 224;
constexpr size_t nb_planes = 3;
cv::Mat planes[nb_planes];
constexpr size_t size = lix * liy * nb_planes;
uint8_t combined_planes[224 * 224 * 3];
planarize(img, combined_planes);
std::vector<uint8_t> combined_planes(size);
planarize(img, combined_planes.data());
// Normalize - supposedly, the keypoint estimator wants keypoints in [0,1]
float real_thing[lix * liy * 3] = {0};
for (size_t i = 0; i < lix * liy * 3; i++) {
std::vector<float> real_thing(size);
for (size_t i = 0; i < size; i++) {
real_thing[i] = (float)combined_planes[i] / 255.0;
}
const OrtApi *g_ort = htv->htd->ort_api;
struct ModelInfo *model = &htv->keypoint_model;
OrtValue *input_tensor = nullptr;
ORT_CHECK(g_ort, g_ort->CreateTensorWithDataAsOrtValue(model->memoryInfo, real_thing, model->input_size_bytes,
model->input_shape.data(), model->input_shape.size(),
ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, &input_tensor));
ORT_CHECK(g_ort, g_ort->CreateTensorWithDataAsOrtValue(
model->memoryInfo, real_thing.data(), model->input_size_bytes, model->input_shape.data(),
model->input_shape.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, &input_tensor));
// Cargo-culted
assert(input_tensor != nullptr);
@ -510,22 +510,24 @@ runHandDetector(struct ht_view *htv, cv::Mat &raw_input)
cv::Mat img;
const int hd_size = 128;
const int inputTensorSize = hd_size * hd_size * 3 * sizeof(float);
constexpr int hd_size = 128;
constexpr size_t nb_planes = 3;
constexpr size_t size = hd_size * hd_size * nb_planes;
constexpr int inputTensorSize = size * sizeof(float);
cv::Matx23f back_from_blackbar = blackbar(raw_input, img, {hd_size, hd_size});
float scale_factor = back_from_blackbar(0, 0); // 960/128
assert(img.isContinuous());
float mean = 128.0f;
float std = 128.0f;
float real_thing[hd_size * hd_size * 3];
constexpr float mean = 128.0f;
constexpr float std = 128.0f;
std::vector<float> real_thing(size);
if (htv->htd->runtime_config.palm_detection_use_mediapipe) {
uint8_t combined_planes[hd_size * hd_size * 3] = {0};
planarize(img, combined_planes);
for (size_t i = 0; i < hd_size * hd_size * 3; i++) {
std::vector<uint8_t> combined_planes(size);
planarize(img, combined_planes.data());
for (size_t i = 0; i < size; i++) {
float val = (float)combined_planes[i];
real_thing[i] = (val - mean) / std;
}
@ -534,7 +536,7 @@ runHandDetector(struct ht_view *htv, cv::Mat &raw_input)
assert(img.isContinuous());
for (size_t i = 0; i < hd_size * hd_size * 3; i++) {
for (size_t i = 0; i < size; i++) {
int val = img.data[i];
real_thing[i] = (val - mean) / std;
@ -548,7 +550,7 @@ runHandDetector(struct ht_view *htv, cv::Mat &raw_input)
OrtValue *input_tensor = nullptr;
ORT_CHECK(g_ort, g_ort->CreateTensorWithDataAsOrtValue(
model_hd->memoryInfo, real_thing, inputTensorSize, model_hd->input_shape.data(),
model_hd->memoryInfo, real_thing.data(), inputTensorSize, model_hd->input_shape.data(),
model_hd->input_shape.size(), ONNX_TENSOR_ELEMENT_DATA_TYPE_FLOAT, &input_tensor));
// Cargo-culted