NonMaxSuppression

对检测框进行非极大值抑制（Non-Maximum Suppression, NMS），用于在目标检测中选取得分较高且重叠度低的边界框，并去除那些重叠度较高的低得分框。其目的是减少冗余边界框，提高目标检测的准确性。

\[ \begin{align}\begin{aligned}\text{对于每个类别 } c, \text{在同一批次下保留置信度较高且 IoU 小于阈值的框:}\\\text{若 } \mathrm{IoU}(box_i, box_j) > \text{iou\_threshold} \Rightarrow \text{抑制 } box_j,\end{aligned}\end{align} \]
\[IoU=\frac{intersec\_area}{(cand.get\_area() + box.get\_area() - intersec\_area)}\]
输入：

  • NmsParam - 输入参数结构体。

  • core_mask(int, 可选) - 核掩码（仅适用于共享存储版本）。

结构体定义：

 1typedef struct {
 2    float* boxes;               // 形状为 (batch, num_boxes, 4)
 3    float* score;               // 形状为 (batch, class, num_boxes)
 4    int* output;                // 输出缓冲区
 5    float* candidates;          // 候选框临时缓冲区
 6    int batch_num;              // 批次数量
 7    int class_num;              // 类别数量
 8    int box_num;                // 每个类别的框数量
 9    int center_point_box;       // 0: corner box, 1: center point box
10    bool simple_out;            // true: 仅输出index, false: 输出 batch, class, index
11    int max_output_per_class;   // 每个类别最大输出框数
12    float iou_threshold;        // IoU 阈值
13    float score_threshold;      // 分数阈值
14} NmsParamFp32;
15typedef struct {
16    int8_t* boxes;               // 形状为 (batch, num_boxes, 4)
17    int8_t* score;               // 形状为 (batch, class, num_boxes)
18    int* output;                // 输出缓冲区
19    float* candidates;          // 候选框临时缓冲区
20    int batch_num;              // 批次数量
21    int class_num;              // 类别数量
22    int box_num;                // 每个类别的框数量
23    int center_point_box;       // 0: corner box, 1: center point box
24    bool simple_out;            // true: 仅输出index, false: 输出 batch, class, index
25    int max_output_per_class;   // 每个类别最大输出框数
26    float iou_threshold;        // IoU 阈值
27    int8_t score_threshold;      // 分数阈值
28} NmsParamInt8;
29typedef struct {
30    float16* boxes;
31    float16* score;
32    int* output;
33    float16* candidates;
34    uint64_t batch_num;
35    uint64_t class_num;
36    uint64_t box_num;
37    uint64_t center_point_box;
38    uint64_t simple_out;
39    uint64_t max_output_per_class;
40    float iou_threshold_bits;
41    float score_threshold_bits;
42} NmsParamFp16;
支持平台：

FT78NE MT7004

备注

  • FT78NE 支持 fp32, int8

  • MT7004 支持 fp16, fp32

  • MT7004中使用结构体作为参数传入。

共享存储版本:

void i8_non_max_suppression_s(NmsParamInt8 *param, int core_mask)
void fp_non_max_suppression_s(NmsParamFp32 *param, int core_mask)
void hp_non_max_suppression_s(NmsParamFp16 *param, int core_mask)

C调用示例：

 1//MT7004示例
 2#include <stdio.h>
 3
 4int main(int argc, char* argv[]) {
 5    float *boxes = (float *)0xA0000000;
 6    float *score = (float *)0xA1000000;
 7    int *output = (int *)0xA2000000;
 8    NmsParamFp32* param = (NmsParamFp32*)0xA3000000;
 9
10    param->boxes = boxes;
11    param->score = score;
12    param->output = output;
13    param->candidates = candidates;
14    param->batch_num = 1;
15    param->class_num = 3;
16    param->box_num = 100;
17    param->center_point_box = 0;
18    param->simple_out = 1;
19    param->max_output_per_class = 50;
20    param->iou_threshold_bits = 0.5;
21    param->score_threshold_bits = 0.1;
22
23    int core_mask = 0xff;
24    fp_non_max_suppression_s(param, core_mask);
25    return 0;
26}

私有存储版本:

void i8_non_max_suppression_p(NmsParamInt8 *param)
void fp_non_max_suppression_p(NmsParamFp32 *param)
void hp_non_max_suppression_p(NmsParamFp16 *param)

C调用示例：

 1//FT78NE示例
 2#include <stdio.h>
 3
 4int main(int argc, char* argv[]) {
 5    float *boxes = (float *)0x1000000;
 6    float *score = (float *)0xA1000000;
 7    int *output = (int *)0xA2000000;
 8    NmsParamFp32* param = (NmsParamFp32*)0xA3000000;
 9
10    param->boxes = boxes;
11    param->score = score;
12    param->output = output;
13    param->candidates = candidates;
14    param->batch_num = 1;
15    param->class_num = 3;
16    param->box_num = 100;
17    param->center_point_box = 0;
18    param->simple_out = 1;
19    param->max_output_per_class = 50;
20    param->iou_threshold_bits = 0.5;
21    param->score_threshold_bits = 0.1;
22
23    fp_non_max_suppression_s(param);
24    return 0;
25}