3808
- 收藏
- 点赞
- 分享
- 举报
hi3516d v300使用yolov3训练数据集仿真正确,板上运行识别不出结果
最近在研究Hi3516D V300运行yolov3,采用Hi3516CV500_SDK_V2.0.2.1
自己训练数据集(3个类别),仿真正常,识别出错,哪位同仁遇到过相同问题或已经成功运行的还望不吝赐教,多谢!
步骤如下
一 板上nnie测试例子已经运行成功;
二 自己训练的数据集,功能、指令仿真都正常,置信度0.99、0.97,类别识别、位置都准确;
>>>>>>> HI_MPI_SVP_NNIE_Forward(Seg-0) FINISH <<<<<<<
n:1, c:24, h:13, w:13
n:1, c:24, h:26, w:26
n:1, c:24, h:52, w:52
u32BoxResultNum: 2
==============================================================================
416 416
test_416x416 0 0.99276584 90.98 98.06 201.46 146.06
test_416x416 1 0.96975088 184.34 129.41 201.01 416.00
==============================================================================
SvpSampleCnnDetYoloV3 end ...
三 修改nnie sample识别不出结果
# ./sample_nnie_main 8
[Level]:Debug,[Func]:SAMPLE_COMM_SVP_CheckSysInit [Line]:81 [Info]:Svp mpi init ok!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3408 [Info]:Yolov3 Load model!
random: crng init done
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3417 [Info]:Yolov3 parameter initialization!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3424 [Info]:Yolov3 start!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3461 [Info]:Yolov3 result:
[Level]:Debug,[Func]:SAMPLE_COMM_SVP_CheckSysExit [Line]:94 [Info]:Svp mpi exit ok!
四 附录
修改如下:
1. sample_nnie.c
void SAMPLE_SVP_NNIE_Yolov3(void)
{
HI_CHAR *pcSrcFile = "./data/nnie_image/rgb_planar/test_416x416.bgr"; // 修改,BGR图片文件
HI_CHAR *pcModelName = "./data/nnie_model/detection/yolov3-voc-test_inst.wk"; // 修改,指令仿真的wk文件
...
}
static HI_S32 SAMPLE_SVP_NNIE_Yolov3_SoftwareInit(SAMPLE_SVP_NNIE_CFG_S* pstCfg,
SAMPLE_SVP_NNIE_PARAM_S *pstNnieParam, SAMPLE_SVP_NNIE_YOLOV3_SOFTWARE_PARAM_S* pstSoftWareParam)
{
...
pstSoftWareParam->u32ClassNum = 3; // 修改,类别为3
pstSoftWareParam->au32GridNumHeight[0] = 13;
pstSoftWareParam->au32GridNumHeight[1] = 26;
pstSoftWareParam->au32GridNumHeight[2] = 52;
pstSoftWareParam->au32GridNumWidth[0] = 13;
pstSoftWareParam->au32GridNumWidth[1] = 26;
pstSoftWareParam->au32GridNumWidth[2] = 52;
...
}
2. sample_model_yolo.h
enum {
SVP_SAMPLE_YOLOV3_SRC_WIDTH = 416,
SVP_SAMPLE_YOLOV3_SRC_HEIGHT = 416,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_82 = 13,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_94 = 26,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_106 = 52,
SVP_SAMPLE_YOLOV3_CHANNLENUM = 24,//255, //修改,3x(5+ClassNum)
SVP_SAMPLE_YOLOV3_PARAMNUM = 8,//85, //修改,(5+ClassNum)
SVP_SAMPLE_YOLOV3_BOXNUM = 3,
SVP_SAMPLE_YOLOV3_CLASSNUM = 3,//80, //修改,类别数量
SVP_SAMPLE_YOLOV3_MAX_BOX_NUM = 10
};
3. sample_svp_nnie_software.h
//修改,类别数+5
#define SAMPLE_SVP_NNIE_YOLOV3_EACH_BBOX_INFER_RESULT_NUM 8//85 /*yolov3 inference result num of each bbox*/
自己训练数据集(3个类别),仿真正常,识别出错,哪位同仁遇到过相同问题或已经成功运行的还望不吝赐教,多谢!
步骤如下
一 板上nnie测试例子已经运行成功;
二 自己训练的数据集,功能、指令仿真都正常,置信度0.99、0.97,类别识别、位置都准确;
>>>>>>> HI_MPI_SVP_NNIE_Forward(Seg-0) FINISH <<<<<<<
n:1, c:24, h:13, w:13
n:1, c:24, h:26, w:26
n:1, c:24, h:52, w:52
u32BoxResultNum: 2
==============================================================================
416 416
test_416x416 0 0.99276584 90.98 98.06 201.46 146.06
test_416x416 1 0.96975088 184.34 129.41 201.01 416.00
==============================================================================
SvpSampleCnnDetYoloV3 end ...
三 修改nnie sample识别不出结果
# ./sample_nnie_main 8
[Level]:Debug,[Func]:SAMPLE_COMM_SVP_CheckSysInit [Line]:81 [Info]:Svp mpi init ok!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3408 [Info]:Yolov3 Load model!
random: crng init done
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3417 [Info]:Yolov3 parameter initialization!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3424 [Info]:Yolov3 start!
[Level]:Info,[Func]:SAMPLE_SVP_NNIE_Yolov3 [Line]:3461 [Info]:Yolov3 result:
[Level]:Debug,[Func]:SAMPLE_COMM_SVP_CheckSysExit [Line]:94 [Info]:Svp mpi exit ok!
四 附录
修改如下:
1. sample_nnie.c
void SAMPLE_SVP_NNIE_Yolov3(void)
{
HI_CHAR *pcSrcFile = "./data/nnie_image/rgb_planar/test_416x416.bgr"; // 修改,BGR图片文件
HI_CHAR *pcModelName = "./data/nnie_model/detection/yolov3-voc-test_inst.wk"; // 修改,指令仿真的wk文件
...
}
static HI_S32 SAMPLE_SVP_NNIE_Yolov3_SoftwareInit(SAMPLE_SVP_NNIE_CFG_S* pstCfg,
SAMPLE_SVP_NNIE_PARAM_S *pstNnieParam, SAMPLE_SVP_NNIE_YOLOV3_SOFTWARE_PARAM_S* pstSoftWareParam)
{
...
pstSoftWareParam->u32ClassNum = 3; // 修改,类别为3
pstSoftWareParam->au32GridNumHeight[0] = 13;
pstSoftWareParam->au32GridNumHeight[1] = 26;
pstSoftWareParam->au32GridNumHeight[2] = 52;
pstSoftWareParam->au32GridNumWidth[0] = 13;
pstSoftWareParam->au32GridNumWidth[1] = 26;
pstSoftWareParam->au32GridNumWidth[2] = 52;
...
}
2. sample_model_yolo.h
enum {
SVP_SAMPLE_YOLOV3_SRC_WIDTH = 416,
SVP_SAMPLE_YOLOV3_SRC_HEIGHT = 416,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_82 = 13,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_94 = 26,
SVP_SAMPLE_YOLOV3_GRIDNUM_CONV_106 = 52,
SVP_SAMPLE_YOLOV3_CHANNLENUM = 24,//255, //修改,3x(5+ClassNum)
SVP_SAMPLE_YOLOV3_PARAMNUM = 8,//85, //修改,(5+ClassNum)
SVP_SAMPLE_YOLOV3_BOXNUM = 3,
SVP_SAMPLE_YOLOV3_CLASSNUM = 3,//80, //修改,类别数量
SVP_SAMPLE_YOLOV3_MAX_BOX_NUM = 10
};
3. sample_svp_nnie_software.h
//修改,类别数+5
#define SAMPLE_SVP_NNIE_YOLOV3_EACH_BBOX_INFER_RESULT_NUM 8//85 /*yolov3 inference result num of each bbox*/
我来回答
回答2个
时间排序
认可量排序
认可0
认可0
或将文件直接拖到这里
悬赏:
E币
网盘
* 网盘链接:
* 提取码:
悬赏:
E币
Markdown 语法
- 加粗**内容**
- 斜体*内容*
- 删除线~~内容~~
- 引用> 引用内容
- 代码`代码`
- 代码块```编程语言↵代码```
- 链接[链接标题](url)
- 无序列表- 内容
- 有序列表1. 内容
- 缩进内容
- 图片![alt](url)
相关问答
-
2020-04-14 22:26:39
-
2021-11-26 13:49:00
-
2021-02-01 15:14:50
-
2020-02-10 21:51:59
-
2020-06-28 23:54:54
-
2021-06-23 09:41:16
-
2019-10-28 16:04:50
-
2020-09-24 09:54:15
-
2017-05-24 19:34:09
-
2020-11-04 15:59:17
-
2016-07-12 14:11:11
-
2018-12-23 11:56:33
-
2022-11-10 10:36:44
-
2019-10-10 14:16:22
-
2019-12-02 18:18:40
-
2016-09-27 15:05:11
-
2019-05-27 20:49:31
-
2017-06-23 15:05:57
-
2020-09-20 23:53:00
无更多相似问答 去提问
点击登录
-- 积分
-- E币
提问
—
收益
—
被采纳
—
我要提问
切换马甲
上一页
下一页
悬赏问答
-
50如何获取vpss chn的图像修改后发送至vo
-
5FPGA通过Bt1120传YUV422数据过来,vi接收不到数据——3516dv500
-
50SS928 运行PQtools 拼接 推到设备里有一半画面会异常
-
53536AV100的sample_vdec输出到CVBS显示
-
10海思板子mpp怎么在vi阶段改变视频数据尺寸
-
10HI3559AV100 多摄像头同步模式
-
9海思ss928单路摄像头vio中加入opencv处理并显示
-
10EB-RV1126-BC-191板子运行自己编码的程序
-
10求HI3519DV500_SDK_V2.0.1.1
-
5有偿求HI3516DV500 + OV5647驱动
举报反馈
举报类型
- 内容涉黄/赌/毒
- 内容侵权/抄袭
- 政治相关
- 涉嫌广告
- 侮辱谩骂
- 其他
详细说明
提醒
你的问题还没有最佳答案,是否结题,结题后将扣除20%的悬赏金
取消
确认
提醒
你的问题还没有最佳答案,是否结题,结题后将根据回答情况扣除相应悬赏金(1回答=1E币)
取消
确认