IntenseWebs/openvino
3
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Fix to extract scores for each class in consideration of background label's id (#10500)

Browse Source 
Signed-off-by: Andrew Kwangwoong Park <andrew.kwangwoong.park@intel.com>
This commit is contained in:
Andrew Kwangwoong Park
2022-02-21 15:35:44 +09:00
committed by GitHub
parent 575ded54a9
commit a7fff7447c
1 changed files with 20 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

35
src/plugins/intel_gpu/src/graph/impls/cpu/detection_output.cpp
View File

@@ -268,7 +268,7 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
const auto& args = instance.argument;
// Per image -> For each label: Pair (score, prior index)
std::vector<std::vector<std::vector<std::pair<float, int>>>> final_detections;
std::vector<std::map<int, std::vector<std::pair<float, int>>>> final_detections;
for (int image = 0; image < num_of_images; ++image) {
const std::vector<std::vector<bounding_box>>& bboxes_per_image = all_bboxes[image];
std::vector<std::vector<std::pair<float, int>>>& conf_per_image = confidences[image];
@@ -323,15 +323,15 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
comp_score_descend<std::pair<int, int>>);
score_index_pairs.resize(args.keep_top_k);
std::vector<std::vector<std::pair<float, int>>> new_indices(args.num_classes);
std::map<int, std::vector<std::pair<float, int>>> new_indices;
for (int j = 0; j < static_cast<int>(score_index_pairs.size()); ++j) {
int label = score_index_pairs[j].second.first;
int idx = score_index_pairs[j].second.second;
new_indices[label].emplace_back(score_index_pairs[j].first, idx);
new_indices[label].push_back(std::make_pair(score_index_pairs[j].first, idx));
}
final_detections.emplace_back(new_indices);
final_detections.push_back(new_indices);
} else {
std::vector<std::vector<std::pair<float, int>>> new_indices(args.num_classes);
std::map<int, std::vector<std::pair<float, int>>> new_indices;
for (auto it = indices.begin(); it != indices.end(); ++it) {
int label = it->first;
const std::vector<int>& labelIndices = it->second;
@@ -340,23 +340,23 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
int idx = labelIndices[j];
for (const auto& s : scores) {
if (s.second == idx) {
new_indices[label].emplace_back(s.first, idx);
new_indices[label].push_back(std::make_pair(s.first, idx));
}
}
}
}
final_detections.emplace_back(new_indices);
final_detections.push_back(new_indices);
}
}
int count = 0;
for (int image = 0; image < num_of_images; ++image) {
const std::vector<std::vector<bounding_box>>& bboxes_per_image = all_bboxes[image];
auto& final_detections_per_image = final_detections[image];
for (int label = 0; label < static_cast<int>(final_detections_per_image.size()); ++label) {
for (auto it = final_detections[image].begin(); it != final_detections[image].end(); ++it) {
int label = it->first;
int loc_label = args.share_location ? 0 : label;
const std::vector<bounding_box>& bboxes = bboxes_per_image[loc_label];
const std::vector<std::pair<float, int>>& label_detections = final_detections_per_image[label];
std::vector<std::pair<float, int>>& label_detections = it->second;
for (std::pair<float, int> score_prior : label_detections) {
out_ptr[count * DETECTION_OUTPUT_ROW_SIZE] = (dtype)static_cast<float>(image);
out_ptr[count * DETECTION_OUTPUT_ROW_SIZE + 1] =
@@ -609,7 +609,7 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
const int num_of_priors,
std::vector<std::vector<std::pair<float, std::pair<int, int>>>>& scoreIndexPairs) {
const int num_classes = instance.argument.num_classes;
const int background_label_id = instance.argument.background_label_id;
const int num_of_images = static_cast<int>(confidences.size());
auto input_confidence = instance.confidence_memory();
const float confidence_threshold = instance.argument.confidence_threshold;
@@ -645,10 +645,14 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
confidence_ptr_float += idx;
__m128 threshold = _mm_load_ps1(&confidence_threshold);
for (int prior = 0; prior < num_of_priors; ++prior) {
int cls = 0;
int idx_start = (background_label_id == 0 ? 1 : 0);
int cls = idx_start;
float max_score = 0;
int max_cls = 0;
for (; cls + 3 < num_classes; cls += 4) {
if ((background_label_id == 0) && (cls == idx_start)) {
confidence_ptr_float += 1;
}
__m128 scores = _mm_loadu_ps(confidence_ptr_float);
confidence_ptr_float += 4;
__m128i mask128 = _mm_castps_si128(_mm_cmpgt_ps(scores, threshold));
@@ -657,9 +661,10 @@ struct detection_output_impl : typed_primitive_impl<detection_output> {
}
int mask = _mm_movemask_ps(_mm_castsi128_ps(mask128));
if (mask & 1) {
label_to_scores[cls + 0].emplace_back(_mm_cvtss_f32(scores), prior);
if (_mm_cvtss_f32(scores) > max_score && cls + 0 != 0) {
max_score = _mm_cvtss_f32(scores); max_cls = cls + 0;
float s = _mm_cvtss_f32(scores);
label_to_scores[cls + 0].emplace_back(s, prior);
if ((cls == idx_start) || (s > max_score)) {
max_score = s; max_cls = cls + 0;
}
}
if (mask & 2) {