I solved face attribute analysis problem in hailo-8L with raspberry pi

General
1

## Problem Description and Resolution Summary

A detailed summary of the issue encountered with the face attribute analysis pipeline and the steps taken to resolve it.

:clipboard: The Problem

The primary issue was a fatal error in the GStreamer pipeline during the face attribute analysis stage, immediately after successful face detection.

  • Error Message: [HailoRT] [error] No tensor with name face_attr_resnet_v1_18_rgbx/fc3
  • Symptoms:
    • Face detection worked correctly.
    • Face attribute analysis (age/gender estimation) failed to run.
    • The GStreamer pipeline would terminate prematurely.

:magnifying_glass_tilted_left: Root Cause Analysis

The investigation revealed a clear mismatch between the AI model and the post-processing library.

  1. Tensor Name Mismatch
  • Actual Model Output: face_attr_resnet_v1_18/fc3
  • TAPPAS Library Expectation: face_attr_resnet_v1_18_rgbx/fc3
  • Discrepancy: The library was hardcoded to look for a tensor name with an unnecessary _rgbx suffix.
  1. Locating the Faulty Code The incorrect tensor name was found hardcoded in the TAPPAS C++ source code.
  • File: face_attributes.cpp:124
  • Incorrect Code:C++face_attributes_postprocess(roi, "face_attr_resnet_v1_18_rgbx/fc3");
  • Correct Code:C++face_attributes_postprocess(roi, "face_attr_resnet_v1_18/fc3");