Hundreds of hours and thousands of cherries later… Hectre’s Spectre for Cherries development team is celebrating as they step into the final stages of this technically challenging project.
After conducting all of the initial image capture, the Spectre team then began development of the Spectre for Cherries detection model. The detection model is what Spectre uses to detect a huge sample size of individual cherries on the top of a bin or bucket. Spectre then takes all of those individual tiny cherries it has detected and sizes up each cherry, and then serves up a size distribution graph for our Spectre for Cherry customers.
That size distribution graph, and the raw data which is also available to Spectre customers via CSV download, is gold!
This data provides valuable insights enabling our cherry customers to:
- Assess the pick early, ensuring the right size profile is being picked and adjusting if necessary (and coaching pickers for improved performance)
- Provide early size data straight off the orchard to the packhouse to enable streamlined packing
- Give sales agents confidence to sell at the earliest possible opportunity because they really know what they’ve got to sell
One of the biggest challenges for the team has been determining whether multiple detection models would be required in order to meet demand from Hectre’s cherry growers and packers.
Some customers are keen to use Spectre for Cherries for small buckets of fruit, whilst others are seeking use on large cherry bins. How to ensure accuracy for both use cases when the volumes are drastically different? Tricky!
Fortunately we have some of the best computer vision talent on the planet working at Hectre and after some fantastic collaboration, the team identified the best way forward and were able to meet the requirements for both small buckets and large bins. (How did we do that? Well now, that’s a Hectre secret!)
A major piece of work then followed called machine learning. When we build a new Spectre model, we need to teach the machine what it’s seeing, so that it can keep that information and use it every time it sees a piece of fruit like that. During development, for each cherry that Spectre detects, we need to assess its size and label it using special data labelling techniques.
It’s a huge piece of work, requiring detailed attention, dedicated focus and a passion for accuracy and quality. At Hectre, we’re fortunate to have awesome team members who have exactly that passion!
Thousands of cherries have now been labelled and end to end testing – capturing images of buckets and bins full of cherries through Spectre, loading them up to the cloud, and receiving fast and accurate size distribution results graphs – has now been completed.
We now enter the final stage of the project – beta testing! This is also known as user testing, or customer validation, and involves providing Spectre for Cherries to a number of select customers for them to put Spectre through its paces, challenge the technology, and identify any user issues. NZ cherry leaders, Cherri Global and CentralPac, have been selected as the first cherry businesses to gain access to Spectre for Cherries. Both are known for their commitment to quality and focus on innovation and they will be undertaking beta testing and grader comparisons during the upcoming NZ cherry harvest, and we’re excited to be working with them.
Cherri Global are implementing rain cover technology to proactively manage the weather risk.
CentralPac use the latest in sorting technologies to ensure the right cherries reach the right market.
During beta testing we’ll be capturing large volumes of cherry images to feed the hungry appetite of Spectre for Cherries. The more images we feed Spectre, the stronger the model becomes, the more fine tuning that can be undertaken, and the greater levels of accuracy we can deliver.
