I am ready to give up - hailo8 26tops. Maybe I am wrong.

General
1

Hello. My first post here. So i installed Hailo8 26 tops on raspberry 5 pi, tested movies detection and id, rtc… OK so they are working… So my biggest question is What Now?
After I read the documentation, the forum, I admit that I don’t understand 99.9% post You wrote here. So, congratulations. What I am doing here?
I\ve checked guides about hailo8 and realized mostly they are not using hailo8 at all. I short, I am lost.
My strategy is to put hailo8 to recycle bin or wait 2 years for more development from the talented hailo team and forum users.
I wish I knew how incompatible and hard to use is hailo8 at the moment with other standards.

2

Hi @kontakt

Thank you for reaching out and sharing your experience. I completely understand how diving into new hardware like the Hailo-8 can feel overwhelming, especially when it’s different from more familiar systems like GPUs.
Deploying AI models on an embedded system can be more challenging than running PyTorch or Tensorflow, so I am here to help clarify things and guide you in the right direction.

First of all - sorry for being direct - I haven’t really understood what difficulties you are experiencing. In your post, you mentioned that you managed to set up the RaspberryPi + Hailo-8L and run a few demos. What comes next depends on what you are trying to achieve in your project, for example:

  • Are you looking to run specific AI models (e.g., object detection, classification)?
  • Are you encountering difficulties with model conversion or inference?
  • Do you have a particular application or problem you’re trying to solve?

Hailo documentation is pretty exhaustive. If you could give some details about what you are trying to do, I can point you to the right documents and examples.

1 Like
3

Thank you for your answer. I appreciate it much. I would like to give you more details about what I am trying to do. I made my first linux debian serwer and I have got 8Gb RAM raspberry pi’s equipped with hailo8 26TOPS.

  1. I am going to make high quality video with existing solutions about utilisation of object detection, classificatio, model conversion, and inference and put it to my website www.gdansk.ai (Gdansk it is a City in Poland, Europe)
  2. The GOAL is to start earn some money. I am going towards all of these technologies you mentioned in your post. I would like to be ready to have a JOB from companies interested in ai hailo possibilities and I would like to be ready to have an offer for companies.
    Example of solutions:
    1/ Project “Monitoring of animals in my city” - when hailo will find/detect a dog, the image photo and time and direction of the dog should be saved on the disk). Goal is to have information about missing animals.
4

Hi @kontakt,

Thanks for clarifying. Since you plan to implement an animal monitoring application, I assume you already have or plan to have a detection model trained for that task.
In any case, you will have to perform these 3 macro-tasks:

1. AI Model Training

Begin by identifying a detection model that aligns with your requirements.

  • The Hailo Model Zoo provides pre-trained detection models, including popular ones like YOLO, which can serve as a starting point. However, if you need better accuracy you may have to retrain the model on your custom dataset. For some models, retrained dockers are available in the Hailo Model Zoo
  • In alternative, you can use your custom model, provided that its operators are compatible with the Hailo Dataflow Compiler (DFC). Check the DFC User Guide to verify the compatibility.
  • Model training does not require Hailo tools, as it is performed in AI frameworks like Tensorflow or PyTorch

2. AI Model Conversion

  • Once you have your trained model, it must be converted into a format compatible with the Hailo-8 accelerator.
  • Use the Hailo Model Zoo or Hailo Dataflow Compiler to convert the model to a Hailo-compatible .hef file.
  • The conversion process involves quantizing the model to optimize performance and power efficiency while maintaining accuracy.

3. Application Development

  • Develop your wildlife detection application to run inference on the Hailo-8/8L accelerator using the HailoRT runtime library.

  • Key components of your application include:

    • Loading the compiled .hef model onto the Hailo-8 device.
    • Capturing video or images from cameras.
    • Running inference with HailoRT to detect dogs in real-time.
    • Implementing any additional logic, such as triggering alerts or recording detections.

  • You can use the Hailo Raspberry Pi5 as a basis for development, or write your own C++/Python code. Some examples are available on GitHub

2 Likes