Skip to content

Commit

Permalink
Implement ailia update
Browse files Browse the repository at this point in the history
  • Loading branch information
@kyakuno
kyakuno committed Nov 1, 2024
1 parent 5306417 commit 305b66f
Showing 1 changed file with 90 additions and 18 deletions.
108 changes: 90 additions & 18 deletions face_detection/retinaface/retinaface.cpp
View file
Original file line number Diff line number Diff line change
Expand Up @@ -381,7 +381,35 @@ vector<int> Nms(const vector<vector<float>>& boxes, const vector<float>& scores,
return keep;
}

bool compare_indices(const std::pair<int, float>& a, const std::pair<int, float>& b) {
return a.second > b.second; // 二番目の要素（スコア）が大きい方を優先
}

std::vector<int> sort_indices_by_score(const std::vector<float>& filtered_score) {
// インデックスとスコアをペアにして保持
std::vector<std::pair<int, float>> indexed_scores;
for (size_t i = 0; i < filtered_score.size(); ++i) {
indexed_scores.emplace_back(i, filtered_score[i]);
}

// スコアに基づいてソート
std::sort(indexed_scores.begin(), indexed_scores.end(), compare_indices);

// ソートされたインデックスを取得
std::vector<int> order;
for (const auto& pair : indexed_scores) {
order.push_back(pair.first);
}

return order;
}

vector<FaceInfo> PostProcess(const vector<float>& box_data, const vector<float>& score_data, const vector<float>& landmark_data, int tex_width, int tex_height) {
PRINT_OUT("PostProcess1");
fflush(stdout);

vector<FaceInfo> results;

vector<int> boxShape = {static_cast<int>(box_data.size() / BOX_DIM), BOX_DIM};
vector<vector<float>> reshaped_box_data = Reshape(box_data, boxShape[1]);
vector<vector<float>> priors = PriorBoxForward(tex_width, tex_height);
Expand All @@ -399,25 +427,58 @@ vector<FaceInfo> PostProcess(const vector<float>& box_data, const vector<float>&
vector<int> landmark_scale = {tex_width, tex_height, tex_width, tex_height, tex_width, tex_height, tex_width, tex_height, tex_width, tex_height};
vector<vector<float>> scaled_landmarks = Scale(landmarks, landmark_scale);

PRINT_OUT("PostProcess2");
fflush(stdout);

vector<int> inds;
for (int i = 0; i < scores.size(); i++) {
if (scores[i] > CONFIDENCE_THRES) {
inds.push_back(i);
}
}

PRINT_OUT("PostProcess3");
fflush(stdout);

PRINT_OUT("scaled_boxes %d", scaled_boxes.size());
PRINT_OUT("scores %d", scores.size());
PRINT_OUT("scaled_landmarks %d", scaled_landmarks.size());
fflush(stdout);

vector<vector<float>> filtered_boxes = Filter(scaled_boxes, inds);
vector<float> filtered_scores = Filter(scores, inds);
vector<vector<float>> filtered_landmarks = Filter(scaled_landmarks, inds);

PRINT_OUT("filtered_boxes %d\n", filtered_boxes.size());
PRINT_OUT("filtered_scores %d\n", filtered_scores.size());
PRINT_OUT("filtered_landmarks %d\n", filtered_landmarks.size());
fflush(stdout);

PRINT_OUT("PostProcess4");
fflush(stdout);

if (filtered_scores.size() == 0){
return results;
}

int top_k = min(TOP_K, static_cast<int>(filtered_scores.size()));
vector<int> order = sort_indices_by_score(filtered_scores);
/*
vector<int> order = [&filtered_scores]() {
vector<int> idx(filtered_scores.size());
iota(idx.begin(), idx.end(), 0);
sort(idx.begin(), idx.end(), [&filtered_scores](int a, int b) { return filtered_scores[a] > filtered_scores[b]; });
return idx;
}();
*/
PRINT_OUT("KeepTopKBeforeNMS filtered_boxes %d %d\n", filtered_boxes.size(), filtered_boxes[0].size());
fflush(stdout);
vector<vector<float>> top_k_boxes = KeepTopKBeforeNMS(filtered_boxes, order, top_k);
PRINT_OUT("KeepTopKBeforeNMS filtered_scores %d\n", filtered_scores.size());
fflush(stdout);
vector<float> top_k_scores = KeepTopKBeforeNMS(filtered_scores, order, top_k);
PRINT_OUT("KeepTopKBeforeNMS filtered_landmarks %d %d\n", filtered_landmarks.size(), filtered_landmarks[0].size());
fflush(stdout);
vector<vector<float>> top_k_landmarks = KeepTopKBeforeNMS(filtered_landmarks, order, top_k);

vector<int> keep = Nms(top_k_boxes, top_k_scores, NMS_THRES);
Expand All @@ -434,7 +495,6 @@ vector<FaceInfo> PostProcess(const vector<float>& box_data, const vector<float>&
}

int keep_top_k = min(KEEP_TOP_K, static_cast<int>(nms_scores.size()));
vector<FaceInfo> results;
for (int i = 0; i < keep_top_k; i++) {
FaceInfo faceInfo;
faceInfo.score = nms_scores[i];
Expand All @@ -450,16 +510,16 @@ vector<FaceInfo> PostProcess(const vector<float>& box_data, const vector<float>&
return results;
}

vector<FaceInfo> Detection(AILIANetwork *ailia, const unsigned char* camera, int stride, int tex_width, int tex_height) {
vector<FaceInfo> Detection(AILIANetwork *ailia, const unsigned char* camera, int tex_width, int tex_height, int channels) {
// Prepare input data
std::vector<float> data(tex_width * tex_height * 3 * 1);

for (int y = 0; y < tex_height; y++) {
for (int x = 0; x < tex_width; x++) {
int idx = (y * stride + x) * 4;
data[(y * tex_width + x) + 2 * tex_width * tex_height] = static_cast<float>(camera[idx + 2] - 123.0f); //R
data[(y * tex_width + x) + 1 * tex_width * tex_height] = static_cast<float>(camera[idx + 1] - 117.0f); //G
data[(y * tex_width + x) + tex_width * tex_height] = static_cast<float>(camera[idx] - 104.0f); //B
int idx = (y * tex_width + x) * channels;
data[(y * tex_width + x) + 2 * tex_width * tex_height] = static_cast<float>(camera[idx + 2]) - 123.0f; //R
data[(y * tex_width + x) + 1 * tex_width * tex_height] = static_cast<float>(camera[idx + 1]) - 117.0f; //G
data[(y * tex_width + x) ] = static_cast<float>(camera[idx + 0]) - 104.0f; //B
}
}

Expand All @@ -469,9 +529,16 @@ vector<FaceInfo> Detection(AILIANetwork *ailia, const unsigned char* camera, in
shape.z = 3;
shape.w = 1;
shape.dim = 4;
ailiaSetInputBlobShape(ailia, &shape, 0, AILIA_SHAPE_VERSION);

ailiaSetInputBlobData(ailia, &data[0], data.size() * sizeof(float), 0);
unsigned int input_idx = 0;
ailiaGetBlobIndexByInputIndex(ailia, &input_idx, 0);

PRINT_OUT("input_idx %d", input_idx);

ailiaSetInputBlobShape(ailia, &shape, input_idx, AILIA_SHAPE_VERSION);
ailiaSetInputBlobData(ailia, &data[0], data.size() * sizeof(float), input_idx);

ailiaUpdate(ailia);

AILIAShape box_shape;
AILIAShape score_shape;
Expand All @@ -492,9 +559,16 @@ vector<FaceInfo> Detection(AILIANetwork *ailia, const unsigned char* camera, in
vector<float> score_data(score_shape.x * score_shape.y * score_shape.z * score_shape.w);
vector<float> landmark_data(landmark_shape.x * landmark_shape.y * landmark_shape.z * landmark_shape.w);

PRINT_OUT("box %d %d %d %d\n", box_shape.x, box_shape.y, box_shape.z, box_shape.w);
PRINT_OUT("score %d %d %d %d\n", score_shape.x, score_shape.y, score_shape.z, score_shape.w);
PRINT_OUT("landmark %d %d %d %d\n", landmark_shape.x, landmark_shape.y, landmark_shape.z, landmark_shape.w);

ailiaGetBlobData(ailia, &box_data[0], box_data.size() * sizeof(float), box_idx);
ailiaGetBlobData(ailia, &score_data[0], score_data.size() * sizeof(float),score_idx);
ailiaGetBlobData(ailia, &landmark_data[0], landmark_data.size() * sizeof(float),landmark_idx);
ailiaGetBlobData(ailia, &score_data[0], score_data.size() * sizeof(float), score_idx);
ailiaGetBlobData(ailia, &landmark_data[0], landmark_data.size() * sizeof(float), landmark_idx);

printf("blob get success");
fflush(stdout);

// Post-processing
vector<FaceInfo> detections = PostProcess(box_data, score_data, landmark_data, tex_width, tex_height);
Expand All @@ -514,9 +588,9 @@ int plot_result_retinaface(std::vector<FaceInfo> info, cv::Mat& img, bool loggin
PRINT_OUT("+ idx=%d\n prob=%.15f\n x=%.15f\n y=%.15f\n w=%.15f\n h=%.15f\n",
i, obj.score, obj.center.first, obj.center.second, obj.width, obj.height);

cv::Point top_left((int)(img.cols*(obj.center.first - obj.width / 2)), (int)(img.rows*(obj.center.second - obj.height / 2)));
cv::Point bottom_right((int)(img.cols*(obj.center.first + obj.width / 2)), (int)(img.rows*(obj.center.second + obj.height / 2)));
cv::Point text_position((int)(img.cols*(obj.center.first - obj.width / 2))+4, (int)(img.rows*(obj.center.second - obj.height / 2)+4));
cv::Point top_left((int)((obj.center.first - obj.width / 2)), (int)((obj.center.second - obj.height / 2)));
cv::Point bottom_right((int)((obj.center.first + obj.width / 2)), (int)((obj.center.second + obj.height / 2)));
cv::Point text_position((int)((obj.center.first - obj.width / 2))+4, (int)((obj.center.second - obj.height / 2)+4));

// update image
cv::Scalar color = hsv_to_rgb(256*((float)i/(float)info.size()), 255, 255);
Expand Down Expand Up @@ -550,8 +624,7 @@ static int recognize_from_image(AILIANetwork* ailia)
PRINT_OUT("BENCHMARK mode\n");
for (int i = 0; i < BENCHMARK_ITERS; i++) {
clock_t start = clock();
results = Detection(ailia, img.data,
img.cols*4, img.cols, img.rows);
results = Detection(ailia, img.data, img.cols, img.rows, img.channels());
clock_t end = clock();
if (status != AILIA_STATUS_SUCCESS) {
PRINT_ERR("ailiaDetectorCompute failed %d\n", status);
Expand All @@ -561,8 +634,7 @@ static int recognize_from_image(AILIANetwork* ailia)
}
}
else {
results = Detection(ailia, img.data,
img.cols*4, img.cols, img.rows);
results = Detection(ailia, img.data, img.cols, img.rows, img.channels());
}

status = plot_result_retinaface(results, img, true);
Expand Down Expand Up @@ -611,7 +683,7 @@ static int recognize_from_video(AILIANetwork* ailia)
cv::cvtColor(resized_img, img, cv::COLOR_BGR2BGRA);

vector<FaceInfo> results;
results = Detection(ailia, img.data, img.cols*4, img.cols, img.rows);
results = Detection(ailia, img.data, img.cols, img.rows, 4);

int status = plot_result_retinaface(results, resized_img, false);
if (status != AILIA_STATUS_SUCCESS) {
Expand Down

0 comments on commit 305b66f

Please sign in to comment.