图片推理结果为空(怎么排查问题出现在哪)

Viewed 50

问题描述


用的demo是k230_sdk\src\reference\ai_poc\face_detection
就加了句打印:printf("results=%d\n",results.size());
在开发板执行:./face_detect_image.sh
case 1./face_detection.elf built at Nov 27 2025 11:13:53
FaceDetection set_input init took 0.081407 ms
FaceDetection set_output_init took 0.380074 ms
FaceDetection pre_process image took 10.0935 ms
FaceDetection run took 4.322 ms
FaceDetection get_output took 0.015037 ms
FaceDetection post_process took 7.00459 ms
results=18
在我的硬件执行:./face_detect_image.sh
case 1./face_detection.elf built at Nov 27 2025 11:13:53
FaceDetection set_input init took 0.082296 ms
FaceDetection set_output_init took 0.384 ms
FaceDetection pre_process image took 10.129 ms
FaceDetection run took 4.33511 ms
FaceDetection get_output took 0.015149 ms
FaceDetection post_process took 7.01463 ms
results=0
为什么我的硬件板子推理结果是0。我对比与开发板的uboot和内核启动信息是一致的。两者烧录的系统固件rtsamrt+linux是一样的,执行的demo也是一致。因为没有报错信息,我也不知道具体问题出现在哪。

硬件板卡


自己的板子(参照韦东山板子修改)

软件版本


拉的最新系统: tag: v2.0

2 Answers

已解决:也不是推理为空,就是不稳定,大部分空,图片识别的结果是固定,但是我们电路输出结果是不稳定,最终修改KPU电压从0.8V到0.9V之后就能稳定识别(但是开发板也是0.8V但是能稳定识别)

图像正常么

摄像头能正常出图,图像正常的。我这里面用的是图片推理,跟图像没关系把。

跟模型没关系,模型推理过程都是一致的,我们这没有你的板子啊

改过阈值么