・ Unique technology delivers fast and accurate CNN learning
・ Fully edge-based image recognition (no cloud or network connection required)
・ Execution speed 2 to 4 times faster than Caffe with only 1/3 of the memory space
・ CNN learning time greatly shortened from several months to one week
・ Optimized for embedded applications
Use cases
・ Image recognition in edge devices
・ Intelligent IoT
・ Antitheft or surveillance devices
・ Machine vision
・ Human behavior recognition
・ Abnormality recognition
・ AR, computer games
Specifications and price
| Kit |
KitA |
KitB
|
KitC |
| Number of recognition targets |
Up to 5 types |
Up to 10 types |
10 types or more |
| Minimum number of images per recognition target |
300
|
300 |
300 |
| Number of times CNN learning performed |
Up to 1 time |
Up to 2 times |
（*） |
Number of times images uploaded |
Up to 1 time
|
Up to 2 times
|
（*） |
| Deliverable |
Image recognition app (supports Android or Linux) |
Image recognition app (supports Android or Linux) |
Image recognition app (*) (supports Android, Linux, or other operating system) |
| Price |
\ 500,000 |
\ 800,000 |
（*） |
* Determined based on user request.
Main flow for use
・ Purchase Kit (A/B/C) according to user request
・ Install dedicated app for Android smartphones
・ Collect or shoot images for deep learning
・ After collecting or shooting sufficient images for each recognition target, upload the images to the specified server
- At least 300 images required for each recognition target, no upper limit
- Can be uploaded from a smartphone app or PC application
・ After about 1 week, an app that can recognize each target is provided as a deliverable from Uncanny Vision
Example of actually using an image recognition app
・ The app provided by Uncanny Vision with machine learning completed is installed.
・ When the recognition target image is shot, the deep learning recognition operation is started.
・ The recognition result and probability are displayed on the screen.
Precautions
・ The recognition target is limited to objects and scenes.
・ The CNN to use for learning is selected by Uncanny Vision.
・ Because the learning process of deep learning starts as soon as the learning images are uploaded,
please collect or shoot sufficient images before uploading them.
・ The app cannot be used after the predetermined number of times of performing learning has been reached.
・ The expected image recognition rate is about 80%. However, because this depends on the quality of the learning images and the number of images provided,
this is not a guaranteed value.
Related Products
 |
Uncanny Vision
Deep-learning library
|
|
Uncanny Vision
Deep-learning library evaluation kit
|
|
Uncanny Vision
Super high-speed computer vision library
for ARM NEON based processors
|
