Opinion
Is the hype around the OpenCV AI kit worth it?
The OpenCV OAK-1 and OAK-D kits are all around the news if you are an aspiring computer vision enthusiasts. But is it worth it?
OpenCV has gradually become one of the best computer vision libraries. It started with the addition of the DNN module in the release of OpenCV 3.3 and now most of its updates are regarding complex modules like the recent addition of YOLOv4 and EfficientDet modules in OpenCV 4.4. Now, it has created further ripples in the industry with the launch of two hardware modules namely OAK-1 and OAK-D with the difference between them being a normal camera and stereo cameras.
I will be writing this article from a student’s point of view and it is an opinion article. You may well have contrasting views so let me know in the comments. So is it worth buying? To find out let’s first review the device and the capabilities it offers.
Features and Capabilities
It is jointly created by OpenCV and Luxonis and comes with its AI-enabled Myriad-X chips that can perform computer vision applications. It is completely open-sourced with MIT-licensed hardware. It supports all the OS with OpenVINO support so it supports Windows, Linux, and Mac-OS without any issues. These devices will come with an OAK-API which is Python and OpenCV compatible. AI processing is done on the hardware itself which means it does not put any extra load on the system it is connecting and does not need any cloud-based services. This also means that the data can remain safe because it is processed locally.
As seen on its Kickstarter page, it can be used to detect and track objects, perform segmentation, stream 4K videos at 30FPS, and also supports custom neural networks. With the OAK-D kit, live depth can be combined with AI which is seen as a cheat code to improve results. They claim that it requires only a 30-second setup and the features are shown below.
Complete List of features as given on the Kickstarter page
- Neural Inference — Object detection, image classification, semantic segmentation, etc.
- Warp/Dewarp — Support for additional lenses for fish-eye applications
- Object Tracking — Up to 20 objects with unique IDs
- AprilTags — Structured navigation (Apriltag is a visual fiducial system, useful for a wide variety of tasks including augmented reality, robotics, and camera calibration)
- H.264 and H.265 encoding (HEVC, 1080P, and 4K video) — 3.125 MB/s for 4K video and a Pi Zero can record 4K/30FPS video with this!
- Feature Tracking — Optical and visual-inertial navigation.
- JPEG encoding — 12MP stills
- Motion estimation — Allows real-time background subtraction
- MJPEG encoding — For easy web streaming, etc.
- Edge detection Harris filtering.
OAK-1 Specific Features
- Automatic Motion-based lossless zooming:
- 12x lossless zoom with 720p output
- 6x lossless zoom with 1080 output
- 1.5x lossless zoom with 4K output
OAK-D Specific Features
- Stereo Depth (Including median filtering) — Extended disparity and subpixel possible for wider dynamic range.
- 3D object localization — Monocular AI with a stereo disparity depth and stereo AI (i.e. stereo neural inference) for small object/feature support
- Object Tracking in 3D space — 3D trajectory in real-time and enables motion statistics in meters.
Camera Specifications
Given below are the specifications of cameras of both the OAK devices.
OAK Color Camera Specifications:
- Image Sensor: IMX378
- Max Framerate: 60fps
- H.265 Framerate: 30fps
- Resolution: 12 MP (4056 x 3040 pixels)
- Field of View: 81 DFOV° — 68.8 HFOV°
- Lens Size: 1/2.3 inch
- AutoFocus: 8 cm — ∞
- F-number: 2.0
OAK-D Stereo Camera Specifications:
- Synchronized Global Shutter
- Image Sensor: OV9282
- Max Framerate: 120fps
- Pixel Size: 3um x 3um
- Resolution: 1280 x 800 pixels
- Field of View: 81 DFOV° — 71.8 HFOV°
- Lens Size: 1/2.3 inch
- Focus (Fixed): 19.6 cm — ∞
- F-number: 2.2
Additional Benefits by the Kickstarter campaign
They were able to reach their goal within the first 20 minutes of the campaign launching! At the time of writing, US$ 544,058 has been raised with 18 days to go. So even if the Kickstarter was to end now then the backers would get a free course teaching them how to use these devices. Along with that, the OAK-D devices will come with an Inertial Measurement Unit (IMU) sensor that measures and reports orientation, velocity, and gravitational forces through the use of accelerometers and gyroscopes and often magnetometers.
So till I now I just talked about the general capabilities of these devices and on these bases, I feel that the hype around these modules is completely justified. Students might feel that a price of 100$ and 150$ for OAK-1 and OAK-D is a little steep but this price is right now at 50% discount and after the Kickstarter ends on 13 August the prices will double. Moreover, there is no other hardware module like this available, and assembling one from separate components means a more bulky and expensive solution. As indicated in this article in PyImageSearch, a lot more articles on using it are coming, so the difficulty in using it won’t be a problem.
These Spatial AI-powered kits offer students and researchers alike to perform real-world projects instead of working on just the databases available and reporting a test accuracy on the test accuracy but never seeing it work in the real world. These cameras offer a pathway to test models in production environments without needing any other external hardware.
So, in conclusion, I feel if one is an aspiring computer vision student, it is certainly worth it. A lot of development is going to happen on it and it might just do that for computer vision what Raspberry Pi did for hobbyist hardware.
