Multi-network inference

Lubos_Omelina · November 27, 2025, 10:42am

I am relatively new user of the AI Hat in Pi 5. I saw the demos for detection and segmentation in the Examples repository. I like it because it abstracts handling of the model. However I want to run detection first and then landmark estimation - for faces and then run the third model on the face images too, and it is not clear to me howI can run 3 models in sequence.

I tried to install Tappas on the pi to have access to the “tappas-multi-network“ script, but the installation fails.

I also succeeded to run my model using HailoRT only and use the stream from OpenCV Camera. I can run the models but I struggle to decode the results of the face detection model just based on this description: Hailo Dataflow Compiler - Profiler Report

Could you advise me whether there is an easier way to run the models without having to decode everything manually on the pi? Is my approach wrong?

omria · November 27, 2025, 5:03pm

Hey @Lubos_Omelina,

Welcome to the Hailo Community!

I’d recommend checking out these two pipelines:

The key thing to look at is the cropper function in the pipeline. Basically, you’ll run your first model, use the cropper to extract regions from the output, and then feed those cropped regions into your second model.

Here’s the cropper implementation that should help:

github.com/hailo-ai/hailo-apps-infra

hailo_apps/hailo_app_python/core/gstreamer/gstreamer_helper_pipelines.py

main

import os

from hailo_apps.hailo_app_python.core.common.defines import (
    GST_VIDEO_SINK,
    TAPPAS_POSTPROC_PATH_DEFAULT,
    TAPPAS_POSTPROC_PATH_KEY,
)


def get_source_type(input_source):
    # This function will return the source type based on the input source
    # return values can be "file", "mipi" or "usb"
    input_source = str(input_source)
    if input_source.startswith("/dev/video"):
        return "usb"
    elif input_source.startswith("rpi"):
        return "rpi"
    elif input_source.startswith("libcamera"):  # Use libcamerasrc element, not suggested
        return "libcamera"
    elif input_source.startswith("0x"):

This file has been truncated. show original

Let me know if you need any clarification!