bmcv_nms_ext
该接口是bmcv_nms接口的广义形式,支持Hard_NMS/Soft_NMS/Adaptive_NMS/SSD_NMS,用于消除网络计算得到过多的物体框,并找到最佳物体框。
处理器型号支持:
该接口支持BM1684/BM1684X。
接口形式:
bm_status_t bmcv_nms_ext(bm_handle_t handle, bm_device_mem_t input_proposal_addr, int proposal_size, float nms_threshold, bm_device_mem_t output_proposal_addr, int topk, float score_threshold, int nms_alg, float sigma, int weighting_method, float * densities, float eta)
参数说明:
bm_handle_t handle
输入参数。bm_handle 句柄。
bm_device_mem_t input_proposal_addr
输入参数。输入物体框数据所在地址,输入物体框数据结构为 face_rect_t,详见下面数据结构说明。需要调用 bm_mem_from_system()将数据地址转化成转化为 bm_device_mem_t 所对应的结构。
int proposal_size
输入参数。物体框个数。
float nms_threshold
输入参数。过滤物体框的阈值,分数小于该阈值的物体框将会被过滤掉。
bm_device_mem_t output_proposal_addr
输出参数。输出物体框数据所在地址,输出物体框数据结构为 nms_proposal_t,详见下面数据结构说明。需要调用 bm_mem_from_system() 将数据地址转化成转化为 bm_device_mem_t 所对应的结构。
int topk
输入参数。当前未使用,为后续可能的的扩展预留的接口。
float score_threshold
输入参数。当使用Soft_NMS或者Adaptive_NMS时,最低的score threshold。当score低于该值时,score所对应的框将被过滤掉。
int nms_alg
输入参数。不同的NMS算法的选择,包括 Hard_NMS/Soft_NMS/Adaptive_NMS/SSD_NMS。
float sigma
输入参数。当使用Soft_NMS或者Adaptive_NMS时,Gaussian re-score函数的参数。
int weighting_method
输入参数。当使用Soft_NMS或者Adaptive_NMS时,re-score函数选项:包括线性权值和Gaussian权值。可选参数:
typedef enum { LINEAR_WEIGHTING = 0, GAUSSIAN_WEIGHTING, MAX_WEIGHTING_TYPE } weighting_method_e;
线性权值表达式如下:
Gaussian权值表达式如下:
上面两个表达式中,\(\mathcal{M}\) 表示当前score最大的物体框,\(b_i\) 表示其他score比 \(\mathcal{M}\) 低的物体框,\(s_i\) 表示其他score比 \(\mathcal{M}\) 低的物体框的score值,\(N_t\) 表示NMS门限,\(\sigma\) 对应本接口的参数 float sigma。
float* densities
输入参数。Adaptive-NMS密度值。
float eta
输入参数。SSD-NMS系数,用于调整iou阈值。
返回值:
BM_SUCCESS: 成功
其他: 失败
代码示例:
#include <assert.h> #include <stdint.h> #include <stdio.h> #include <algorithm> #include <functional> #include <iostream> #include <memory> #include <set> #include <string> #include <vector> #include <math.h> #include "bmcv_api.h" #include "bmcv_internal.h" #include "bmcv_common_bm1684.h" #define MAX_PROPOSAL_NUM (65535) typedef float bm_nms_data_type_t; typedef struct { float x1; float y1; float x2; float y2; float score; } face_rect_t; typedef struct nms_proposal { int size; face_rect_t face_rect[MAX_PROPOSAL_NUM]; int capacity; face_rect_t *begin; face_rect_t *end; } nms_proposal_t; typedef enum { LINEAR_WEIGHTING = 0, GAUSSIAN_WEIGHTING, MAX_WEIGHTING_TYPE } weighting_method_e; template <typename data_type> static bool generate_random_buf(std::vector<data_type> &random_buffer, int random_min, int random_max, int scale) { for (int i = 0; i < scale; i++) { data_type data_val = (data_type)( random_min + (((float)((random_max - random_min) * i)) / scale)); random_buffer.push_back(data_val); } std::random_shuffle(random_buffer.begin(), random_buffer.end()); return false; } int main(int argc, char *argv[]) { unsigned int seed1 = 100; bm_nms_data_type_t nms_threshold = 0.22; bm_nms_data_type_t nms_score_threshold = 0.22; bm_nms_data_type_t sigma = 0.4; int proposal_size = 500; int rand_loop_num = 10; int weighting_method = GAUSSIAN_WEIGHTING; std::function<float(float, float)> weighting_func; int nms_type = SOFT_NMS; // ADAPTIVE NMS / HARD NMS / SOFT NMS const int soft_nms_total_types = MAX_NMS_TYPE - HARD_NMS - 1; for (int rand_loop_idx = 0;rand_loop_idx < (rand_loop_num * soft_nms_total_types);rand_loop_idx++) { for (int rand_mode = 0; rand_mode < MAX_RAND_MODE; rand_mode++) { std::shared_ptr<Blob<face_rect_t>> proposal_rand = std::make_shared<Blob<face_rect_t>>(MAX_PROPOSAL_NUM); std::shared_ptr<nms_proposal_t> output_proposal = std::make_shared<nms_proposal_t>(); std::vector<face_rect_t> proposals_ref; std::vector<face_rect_t> nms_proposal; std::vector<bm_nms_data_type_t> score_random_buf; std::vector<bm_nms_data_type_t> density_vec; std::shared_ptr<Blob<float>> densities = std::make_shared<Blob<float>>(proposal_size); generate_random_buf<bm_nms_data_type_t>( score_random_buf, 0, 1, 10000); face_rect_t *proposal_rand_ptr = proposal_rand.get()->data; float eta = ((float)(rand() % 10)) / 10; for (int32_t i = 0; i < proposal_size; i++) { proposal_rand_ptr[i].x1 = ((bm_nms_data_type_t)(rand() % 100)) / 10; proposal_rand_ptr[i].x2 = proposal_rand_ptr[i].x1 + ((bm_nms_data_type_t)(rand() % 100)) / 10; proposal_rand_ptr[i].y1 = ((bm_nms_data_type_t)(rand() % 100)) / 10; proposal_rand_ptr[i].y2 = proposal_rand_ptr[i].y1 + ((bm_nms_data_type_t)(rand() % 100)) / 10; proposal_rand_ptr[i].score = score_random_buf[i]; proposals_ref.push_back(proposal_rand_ptr[i]); densities.get()->data[i] = ((float)(rand() % 100)) / 100; } assert(proposal_size <= MAX_PROPOSAL_NUM); if (weighting_method == LINEAR_WEIGHTING) { weighting_func = linear_weighting; } else if (weighting_method == GAUSSIAN_WEIGHTING) { weighting_func = gaussian_weighting; } else { std::cout << "weighting_method error: " << weighting_method << std::endl; } bmcv_nms_ext(handle, bm_mem_from_system(proposal_rand.get()->data), proposal_size, nms_threshold, bm_mem_from_system(output_proposal.get()), 1, nms_score_threshold, nms_type, sigma, weighting_method, densities.get()->data, eta); } } return 0; }
注意事项:
该 api 可输入的最大 proposal 数为 1024。