Yes, we made it! We built the very first apple-size-detector using a pre-trained model (involving the COCO dataset), that already knew how to recognize simple objects.
How did we achieve this?
We used the pre-trained model to detect individual apples and draw masks and bounding boxes around them (see output images below). The geometric information of the masks and bounding boxes can now be translated into real-world dimensions such as millimeters and inches.
Initial apple detector: first output
Initial apple detector: second output
Initial apple detector: third output
Initial apple detector: fourth output
The initial results proved highly accurate once the system had been optimised. However, due to the nature of the pre-trained data-set, the apple detector was identifying the entire top layer of apples as an ‘apple’, in addition to individual apples. Additionally, more detections were made for cropped or close up apple images as compared with images taken from far away. So, there is still plenty of room for improvement, but we’re stoked how quickly we got the first usable results. This proves we’re getting closer to bringing computer vision to our growers.
Watch this space, as we will let you know about the latest developments very soon!