The Best and Worst Cameras for Computer Vision on the Edge
Growing a computer vision startup, I have tested camera after camera to find the best for computation at the edge.
The rise of computer vision has sparked a revolution in how we interact with the world around us, but processing all that visual data in the cloud presents challenges. Transmission of the images is constrained by the size of the files and the amount of power available. This is where edge computing cameras come in, bringing intelligence directly to the source of the data.
If anyone wants to try AI/ML inference at the edge, these are the best and not-so-great cameras to use and why.
PROTOTYPING:
High-Performance Powerhouses:
NVIDIA Jetson Nano:
This modular setup pairs a powerful Jetson Nano carrier board with various MIPI CSI camera modules. It’s a developer’s playground, offering unparalleled customizability and processing power for complex AI tasks like object recognition and tracking. But be prepared for assembly and coding — this is NOT for the faint of heart.
- Pros: Nvidia is skyrocketing. New models are being uploaded within TAO Toolkit almost weekly at this point. Less building means faster deployment along with a large developer community.
- Cons: I may get in trouble for saying this, but working in the Nvidia ecosystem is a nightmare. Configure issues are all too common. Experience with Linux and understanding the Nvidia Architecture is necessary for building and deployment. Price is something to consider with the Jetson, especially if you need to upgrade to a Xavier.
I s*** you not — I spent nearly 5 hours trying to run a preconfigured model in the TAO toolkit. There were configuration issues at every step, and discovered nothing was wrong. Nvidia had not yet updated its software to run with the latest Linux distribution.
Coral Dev Kit:
Featuring Google’s Edge TPU (Tensor Processing Unit), this kit offers user-friendly plug-and-play functionality with USB cameras. Ideal for simple AI tasks like object detection or facial recognition, it’s a prime choice for beginners and rapid prototyping.
- Pros: Coral is wicked cheap. Coming in at only $129 for their prototyping board, this is hard to beat for a fully functional edge device with model processing power.
- Cons: Tensorflow Lite is the software of choice. This is understandable since Google owns both Coral and maintains Tensorflow; however, if more powerful models (i.e. something you build in Pytorch, a full Tensorflow model, LLama2, etc.) are needed, this board will have challenges.
POCKET SIZE PORTABILITY
Insta360 ONE X2:
This 360° camera packs a punch in a compact package. Its high-resolution sensor captures immersive footage, while its built-in FlowState stabilization keeps things smooth. Limited-edge AI features like object tracking and horizon leveling add extra smarts, making it perfect for on-the-go creators.
- Pros: This little camera produces awe-inspiring images. X2 has been making its way around youtube vloggers and extreme sports junkies, but Insta360 has an opportunity to collect some data at the edge. The setup for this camera is relatively simple.
- Cons: Aside from the tiny screen on the camera, if building models to use on the edge is of interest, Insta360 requires approval to use their SDK (this process is easy and nothing to worry about). This camera is better for data collection rather than edge deployment.
GoPro HERO10 Black:
Action-ready and rugged, the HERO10 Black boasts improved image processing and HyperSmooth 4.0 for buttery-smooth videos. While its edge AI capabilities are still nascent, it’s a reliable choice for capturing high-quality footage in demanding environments.
- Pros: At $250, the price for this camera is decent, considering that it can be thrown out of a plane and submerged underwater.
- Cons: Uploading custom models is nearly impossible. Limitations on uploading the images/videos taken on the device are restrictive. There are three ways to get media off the camera: The Quik app, USB, or their cable — not great (I’m guessing this is because the 5k media is too large to transmit).
CONSTANT POWER AND INTERNET
Smart Home and IP Cameras:
Reolink rlc-811a:
This PoE camera packs a punch with built-in AI for object detection, person recognition, and activity zones. It seamlessly integrates into your smart home, sending alerts and auto-recording moments.
- Pros: This PoE camera has a sister version that is wifi enabled. I prefer the PoE version because tapping the RTSP feed is possible with a physical connection. Although I have yet to find a way to add my AI models to the camera, connecting to the feed is almost impossible with all other brands without their permission. The price of this camera is attractive as well — $95.
- Cons: Connecting via PoE is a consideration. These cameras can be power intensive because they were built to be stationary and attached to power & internet. In addition to the camera, an NVR and computation board to run models is also needed. Maybe a pairing with Coral?
Wyze Cam Pro v3:
Affordable and versatile, the Wyze Cam Pro offers on-device object and person detection, making it a budget-friendly option for entry-level computer vision. Cloud storage and subscription options unlock advanced features like facial recognition and package detection.
- Pros: This camera is the cheapest option I’ve discovered to play around with edge AI. The starting point is ~$50 and has out-of-the-box functionality.
- Cons: There is no customization. These cameras are made for out-of-the-box functionality, not tinkering. Similar limitations in other home cameras include data used in the app, limited camera resolution, and not made for outdoor use.
***My TOP PICK***:
Luxonis Oak-D Pro:
This is the all-in-one camera we were looking for. Capable of powerful edge computation, the Oak-D Pro is also fully integrated with a high-powered camera. The cherry on top is that custom machine-learning models can run and deploy.
- Pros: This camera is powerful enough to host and run custom machine-learning models. Additionally, the out-of-the-box models showcase decent use cases. Infrared also allows for depth detection, which would be difficult otherwise.
- Cons: Familiarity with PyTorch is a requirement (unlike the Coral board). If the model can reformat to Onnx format, then this is a non-issue. My only gripe about the Oak-D Pro is that it’s not made for outdoor use, although the temperature is rated.
OLDIE BUT A GOODIE:
Logitech C920x HD Pro:
This is the peak of easy-to-use USB-connected cameras. While not AI-enabled, this camera has been around for a while and directly connects to a board capable of running edge computation.
- Pros: The price of this camera is surprisingly similar to the Wyze Cam Pro. Affordability and ease of use are the name of the game for this camera. Just plug and play.
- Cons: There is no AI capability. This is just a camera — but you can’t beat the classics.
Finding the Perfect Fit:
Ultimately, the best edge computation camera depends on your specific needs. Consider factors like processing power, resolution, portability, connectivity, and future-proofing potential.
I hope my experiences working with these cameras and having tested them for outdoor use, edge deployment, and easy configuration are helpful for anyone trying to discover more about computer vision.
