| 1 |
// |
| 2 |
// Copyright © 2020 Arm Ltd. All rights reserved. |
| 3 |
// SPDX-License-Identifier: MIT |
| 4 |
// |
| 5 |
|
| 6 |
|
| 7 |
#include "NMS.hpp" |
| 8 |
|
| 9 |
#include <cmath> |
| 10 |
#include <algorithm> |
| 11 |
#include <cstddef> |
| 12 |
#include <numeric> |
| 13 |
#include <ostream> |
| 14 |
|
| 15 |
namespace yolov3 { |
| 16 |
namespace { |
| 17 |
/** Number of elements needed to represent a box */ |
| 18 |
constexpr int box_elements = 4; |
| 19 |
/** Number of elements needed to represent a confidence factor */ |
| 20 |
constexpr int confidence_elements = 1; |
| 21 |
|
| 22 |
/** Calculate Intersection Over Union of two boxes |
| 23 |
* |
| 24 |
* @param[in] box1 First box |
| 25 |
* @param[in] box2 Second box |
| 26 |
* |
| 27 |
* @return The IoU of the two boxes |
| 28 |
*/ |
| 29 |
float iou(const Box& box1, const Box& box2) |
| 30 |
{ |
| 31 |
const float area1 = (box1.xmax - box1.xmin) * (box1.ymax - box1.ymin); |
| 32 |
const float area2 = (box2.xmax - box2.xmin) * (box2.ymax - box2.ymin); |
| 33 |
float overlap; |
| 34 |
if (area1 <= 0 || area2 <= 0) |
| 35 |
{ |
| 36 |
overlap = 0.0f; |
| 37 |
} |
| 38 |
else |
| 39 |
{ |
| 40 |
const auto y_min_intersection = std::max<float>(box1.ymin, box2.ymin); |
| 41 |
const auto x_min_intersection = std::max<float>(box1.xmin, box2.xmin); |
| 42 |
const auto y_max_intersection = std::min<float>(box1.ymax, box2.ymax); |
| 43 |
const auto x_max_intersection = std::min<float>(box1.xmax, box2.xmax); |
| 44 |
const auto area_intersection = |
| 45 |
std::max<float>(y_max_intersection - y_min_intersection, 0.0f) * |
| 46 |
std::max<float>(x_max_intersection - x_min_intersection, 0.0f); |
| 47 |
overlap = area_intersection / (area1 + area2 - area_intersection); |
| 48 |
} |
| 49 |
return overlap; |
| 50 |
} |
| 51 |
|
| 52 |
std::vector<Detection> convert_to_detections(const NMSConfig& config, |
| 53 |
const std::vector<float>& detected_boxes) |
| 54 |
{ |
| 55 |
const size_t element_step = static_cast<size_t>( |
| 56 |
box_elements + confidence_elements + config.num_classes); |
| 57 |
std::vector<Detection> detections; |
| 58 |
|
| 59 |
for (unsigned int i = 0; i < config.num_boxes; ++i) |
| 60 |
{ |
| 61 |
const float* cur_box = &detected_boxes[i * element_step]; |
| 62 |
if (cur_box[4] > config.confidence_threshold) |
| 63 |
{ |
| 64 |
Detection det; |
| 65 |
det.box = {cur_box[0], cur_box[0] + cur_box[2], cur_box[1], |
| 66 |
cur_box[1] + cur_box[3]}; |
| 67 |
det.confidence = cur_box[4]; |
| 68 |
det.classes.resize(static_cast<size_t>(config.num_classes), 0); |
| 69 |
for (unsigned int c = 0; c < config.num_classes; ++c) |
| 70 |
{ |
| 71 |
const float class_prob = det.confidence * cur_box[5 + c]; |
| 72 |
if (class_prob > config.confidence_threshold) |
| 73 |
{ |
| 74 |
det.classes[c] = class_prob; |
| 75 |
} |
| 76 |
} |
| 77 |
detections.emplace_back(std::move(det)); |
| 78 |
} |
| 79 |
} |
| 80 |
return detections; |
| 81 |
} |
| 82 |
} // namespace |
| 83 |
|
| 84 |
bool compare_detection(const yolov3::Detection& detection, |
| 85 |
const std::vector<float>& expected) |
| 86 |
{ |
| 87 |
float tolerance = 0.001f; |
| 88 |
return (std::fabs(detection.classes[0] - expected[0]) < tolerance && |
| 89 |
std::fabs(detection.box.xmin - expected[1]) < tolerance && |
| 90 |
std::fabs(detection.box.ymin - expected[2]) < tolerance && |
| 91 |
std::fabs(detection.box.xmax - expected[3]) < tolerance && |
| 92 |
std::fabs(detection.box.ymax - expected[4]) < tolerance && |
| 93 |
std::fabs(detection.confidence - expected[5]) < tolerance ); |
| 94 |
} |
| 95 |
|
| 96 |
void print_detection(std::ostream& os, |
| 97 |
const std::vector<Detection>& detections) |
| 98 |
{ |
| 99 |
for (const auto& detection : detections) |
| 100 |
{ |
| 101 |
for (unsigned int c = 0; c < detection.classes.size(); ++c) |
| 102 |
{ |
| 103 |
if (detection.classes[c] != 0.0f) |
| 104 |
{ |
| 105 |
os << c << " " << detection.classes[c] << " " << detection.box.xmin |
| 106 |
<< " " << detection.box.ymin << " " << detection.box.xmax << " " |
| 107 |
<< detection.box.ymax << std::endl; |
| 108 |
} |
| 109 |
} |
| 110 |
} |
| 111 |
} |
| 112 |
|
| 113 |
std::vector<Detection> nms(const NMSConfig& config, |
| 114 |
const std::vector<float>& detected_boxes) { |
| 115 |
// Get detections that comply with the expected confidence threshold |
| 116 |
std::vector<Detection> detections = |
| 117 |
convert_to_detections(config, detected_boxes); |
| 118 |
|
| 119 |
const unsigned int num_detections = static_cast<unsigned int>(detections.size()); |
| 120 |
for (unsigned int c = 0; c < config.num_classes; ++c) |
| 121 |
{ |
| 122 |
// Sort classes |
| 123 |
std::sort(detections.begin(), detections.begin() + static_cast<std::ptrdiff_t>(num_detections), |
| 124 |
[c](Detection& detection1, Detection& detection2) |
| 125 |
{ |
| 126 |
return (detection1.classes[c] - detection2.classes[c]) > 0; |
| 127 |
}); |
| 128 |
// Clear detections with high IoU |
| 129 |
for (unsigned int d = 0; d < num_detections; ++d) |
| 130 |
{ |
| 131 |
// Check if class is already cleared/invalidated |
| 132 |
if (detections[d].classes[c] == 0.f) |
| 133 |
{ |
| 134 |
continue; |
| 135 |
} |
| 136 |
|
| 137 |
// Filter out boxes on IoU threshold |
| 138 |
const Box& box1 = detections[d].box; |
| 139 |
for (unsigned int b = d + 1; b < num_detections; ++b) |
| 140 |
{ |
| 141 |
const Box& box2 = detections[b].box; |
| 142 |
if (iou(box1, box2) > config.iou_threshold) |
| 143 |
{ |
| 144 |
detections[b].classes[c] = 0.f; |
| 145 |
} |
| 146 |
} |
| 147 |
} |
| 148 |
} |
| 149 |
return detections; |
| 150 |
} |
| 151 |
} // namespace yolov3 |