什么是NMS
NMS（Non-Maximum Suppression，非极大值抑制）是目标检测领域中最常用的后处理算法之一。它的作用是消除冗余的检测框，保留最准确的检测结果。
算法原理
- 对所有检测框按置信度分数从高到低排序
- 选择置信度最高的检测框，将其加入最终结果
- 计算该框与其余所有框的IoU（Intersection over Union）
- 删除IoU超过阈值的框
- 重复步骤2-4，直到所有框都被处理
IoU计算
IoU（Intersection over Union）是两个框的交集面积与并集面积的比值：
code
IoU = 交集面积 / 并集面积
C++ 实现
cpp
#include <vector>
#include <algorithm>
struct BBox {
float x1, y1, x2, y2;
float score;
};
float computeIoU(const BBox& a, const BBox& b) {
float interX1 = std::max(a.x1, b.x1);
float interY1 = std::max(a.y1, b.y1);
float interX2 = std::min(a.x2, b.x2);
float interY2 = std::min(a.y2, b.y2);
float interArea = std::max(0.0f, interX2 - interX1) * std::max(0.0f, interY2 - interY1);
float areaA = (a.x2 - a.x1) * (a.y2 - a.y1);
float areaB = (b.x2 - b.x1) * (b.y2 - b.y1);
return interArea / (areaA + areaB - interArea);
}
std::vector<BBox> nms(std::vector<BBox>& boxes, float threshold) {
std::sort(boxes.begin(), boxes.end(), [](const BBox& a, const BBox& b) {
return a.score > b.score;
});
std::vector<BBox> result;
std::vector<bool> suppressed(boxes.size(), false);
for (size_t i = 0; i < boxes.size(); ++i) {
if (suppressed[i]) continue;
result.push_back(boxes[i]);
for (size_t j = i + 1; j < boxes.size(); ++j) {
if (suppressed[j]) continue;
float iou = computeIoU(boxes[i], boxes[j]);
if (iou > threshold) {
suppressed[j] = true;
}
}
}
return result;
}
Java 实现
java
import java.util.ArrayList;
import java.util.Collections;
import java.util.Comparator;
import java.util.List;
public class NMS {
static class BBox {
float x1, y1, x2, y2;
float score;
public BBox(float x1, float y1, float x2, float y2, float score) {
this.x1 = x1; this.y1 = y1;
this.x2 = x2; this.y2 = y2;
this.score = score;
}
}
public static float computeIoU(BBox a, BBox b) {
float interX1 = Math.max(a.x1, b.x1);
float interY1 = Math.max(a.y1, b.y1);
float interX2 = Math.min(a.x2, b.x2);
float interY2 = Math.min(a.y2, b.y2);
float interArea = Math.max(0, interX2 - interX1) * Math.max(0, interY2 - interY1);
float areaA = (a.x2 - a.x1) * (a.y2 - a.y1);
float areaB = (b.x2 - b.x1) * (b.y2 - b.y1);
return interArea / (areaA + areaB - interArea);
}
public static List<BBox> nms(List<BBox> boxes, float threshold) {
Collections.sort(boxes, (a, b) -> Float.compare(b.score, a.score));
List<BBox> result = new ArrayList<>();
boolean[] suppressed = new boolean[boxes.size()];
for (int i = 0; i < boxes.size(); i++) {
if (suppressed[i]) continue;
result.add(boxes.get(i));
for (int j = i + 1; j < boxes.size(); j++) {
if (suppressed[j]) continue;
float iou = computeIoU(boxes.get(i), boxes.get(j));
if (iou > threshold) suppressed[j] = true;
}
}
return result;
}
}
应用场景
- YOLO系列：YOLOv3、YOLOv4、YOLOv5等
- SSD：Single Shot MultiBox Detector
- Faster R-CNN：Region-based CNN
- 目标跟踪：多目标跟踪中的检测框筛选
参数选择
- IoU阈值：通常设置为0.5或0.3
- 置信度阈值：根据实际需求调整
总结
NMS是目标检测流程中不可或缺的一环，理解其原理并能够手动实现对于深入理解目标检测算法非常有帮助。
评论 (0)