Operation
The NE301 creates its own Wi-Fi access point, allowing you to set it up and connect it to your main Wi-Fi network. Through its web interface, you can view a live video preview, switch between models, upload your own YOLOv8 models, and tweak thresholds and output formats.
Software-wise
The NE301 is an edge AI camera built around the STM32N6 microcontroller. The CamThink team has developed an open source-based web GUI for controlling the camera operations. The product is designed to offer developers, AI enthusiasts, and smart home fans an affordable, low-power AI camera capable of running local AI models based on YOLOv8 (You Only Look Once) which is also open source based.
From our hands-on experience, it works quite well and offers advantages over running object detection AI models through a cloud server. Running local models is much faster, meaning you can enjoy lower latency time. This device can also be used as a security surveillance camera to monitor people and environments, so the only limit is your imagination.
The NE301 comes with CamThink management software pre-installed on its 128MB HyperFlash. To access it, you will need to power the device and connect to it via Wi-Fi as an access point. Once connected, just open a web browser and enter the device’s gateway IP address: http://192.168.10.10. From there, the web interface lets you easily control camera functions and deploy AI models.
The device main operating settings can be grouped into the following categories:
- Load a model (factory or custom)
- Run inference on incoming data (e.g., camera frames)
- Tune thresholds to optimize detection behavior
- Send results upstream via MQTT for monitoring or automation
For example, the Object detection AI identifies and recognizes the objects it sees. So, what it does is:
- Detects people, vehicles, animals, packages, etc.
- Draws bounding boxes around detected objects.
- Classifies objects (e.g., “Person”, “Car”, “Dog”).
- Tracks objects as they move across frames.
Training AI Models
You will need to train your AI models. This is part of the learning process, like teaching a baby its first steps. When you first turn on the NE301 device, it’s worth noting that it won’t automatically successfully recognizes objects or be ready to use straight out of the box. Training AI models for the NeoEyes NE301 is done with CamThink’s AI Tool Stack—a complete pipeline for data collection, annotation, training, quantization, and deploying models to the camera. The NE301 doesn’t train models onboard; instead, you train them on a computer or in the cloud and then deploy them to the camera.
What are the steps?
How to Train AI Models – A Complete Summary of the Workflow
The AI Tool Stack can be installed on a Linux-based computer or even a Mini PC. While an NVIDIA GPU is recommended for better performance, it’s not required. The table below outlines the workflow steps for training and deploying an AI model using the NE301.
| Stage | Tool | What Happens |
|---|---|---|
| 1. Data Collection | NE301 Camera | Capture images/events for training |
| 2. Annotation | AI Tool Stack | Label objects/actions in images |
| 3. Training | AI Tool Stack + Ultralytics | Train YOLO‑based model on PC/cloud |
| 4. Quantization | AI Tool Stack | Convert model to INT8 for STM32N6 |
| 5. Deployment | NE301 Web UI | Upload model and run inference on camera |
We gave a few trained models a spin to see how they performed.
The company kindly sent us a few pre-trained YOLOv8 models to try out. The first one featured object detection, mainly for humans, along with distance approximation to estimate range, which worked rather well. The second model was designed to detect various types of objects captured by the camera but ended up performing poorly, likely due to insufficient AI training or human error, so it’s understandable.
Overall, the web-based GUI is rather easy to operate and offers a wide selection of features and settings to explore. For more accurate AI-based object detection, you can connect the NE301 to the OpenAI API. While this will likely boost accuracy, it could also lead to higher latency or slower response times, since the server has to send and receive data packets over the internet.
NE301 management software
