https://github.com/nickoala/edgetpu-on-pi

Raspberry Pi Supplement to Coral Edge TPU Demo
https://github.com/nickoala/edgetpu-on-pi

coral edgetpu raspberry-pi

Last synced: about 1 month ago
JSON representation

Raspberry Pi Supplement to Coral Edge TPU Demo

• Host: GitHub
• URL: https://github.com/nickoala/edgetpu-on-pi
• Owner: nickoala
• License: apache-2.0
• Archived: true
• Created: 2019-03-17T03:40:43.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2019-12-22T14:35:02.000Z (over 4 years ago)
• Last Synced: 2024-04-29T12:35:35.139Z (4 months ago)
• Topics: coral, edgetpu, raspberry-pi
• Language: Python
• Size: 17.6 KB
• Stars: 10
• Watchers: 4
• Forks: 3
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE.txt

Awesome Lists containing this project

README

        
# Raspberry Pi Supplement to Coral Edge TPU Demo

Additional sample code to run **image classification** and **object detection**

on a [Coral USB accelerator](https://coral.ai/products/accelerator) from a

Raspberry Pi, filling some gaps in the [official

demos](https://github.com/google-coral/edgetpu/tree/master/examples):

- I would like to see *object detection working on a video stream* where

  rectangles are drawn on the preview to indicate object locations.

- Because Pi Camera's frame *width* is restricted to multiples of 32 and

  *height* to multiples of 16, it cannot match some models' required input size.

  For example, if you tell Pi Camera to resize to Inception V3's required

  299x299, it actually rounds up to 320x304. The captured image has to be

  cropped before passed to Edge TPU.

Resulting files are:

1. `classify_capture.py`: Original official demo, kept for reference.

2. `classify_capture_crop.py`: This preemptively checks the model input size

against Pi Camera's capable frame size. The image is cropped, if necessary,

before passed to Edge TPU.

3. `detect_capture_crop.py`: Use the `Annotator` class borrowed from AIY Vision

Kit to draw bounding boxes around detected objects. It also crops images if

camera's frame size cannot match model input size.