Create

Sign in to ReadmeX

or

How we handle your data: Privacy Policy

ReadmeX

Your AI reading companion

What would you like to catch up on?

Pick a briefing, or ask your own question.

Your community and people briefings stay personal to you.

Story

Debugging YOLOv5x on Journey 6M with hrt_model_exec

AI summary

A report documents a debugging workflow for the yolov5x_640x640_nv12 model on a Journey 6M development board using hrt_model_exec. The author separates NV12 input tensors, three raw HBM outputs, and YOLO post-processing, showing that valid inputs and finite, nonzero outputs do not guarantee sensible detection boxes. Changing the confidence threshold from 0.25 to 0.90 reduced the same test image's detections from six to one while the NMS threshold remained unchanged.

In-depth insight

The report offers a reproducible way to distinguish model-output problems from post-processing issues such as class mapping, confidence filtering, NMS, resizing, and coordinate restoration when edge-AI detections look wrong.

在 征程6M 上把 YOLOv5x 部署起来之后,比起推理性能,更容易遇到的是这样一个问题:

hrt_model_exec infer 已经正常结束,为什么最终检测框还是看起来“不对”? 很多时候,问题在于我们把三件不同的事混在了一起:

输入图片 → 实际送入 HBM 的张量 → HBM 原始输出 → YOLO 后处理 → 检测框

hrt_model_exec 很适合把前两段和中间的模型输出拿出来检查;检测框、类别名、置信度阈值和 NMS,则属于后处理链路。下面用 征程6M 上的 yolov5x_640x640_nv12.hbm 做一次完整拆分。

实验环境与输入

本次使用 征程6M 开发板,实际运行时日志显示:

#Journey 6M#hrt_model_exec#YOLOv5x#yolov5x_640x640_nv12#NV12#NMS

Original · 4 min readRead the original
Group chat

No comments yet. Start the conversation.