Point-and-Shoot Cluster Counting in myEV
- Nicholas Gunner
As part of our effort to make computer vision accessible to viticulturists, we trained several flower cluster object detection models. These models can be used to automatically detect and count visible clusters in images.
You may be wondering, ‘why create several models to count grape clusters?’ Good question. With so many varieties, locations, trellis systems, etc. it is hard to accurately represent all variations of clusters in a single model. We also tested model training at night with artificial lights and during the day. We captured images in New York, California, and Washington. As we refine our models, we may add more options to choose from or we may consolidate some of them. Either way, you will have to consider which model will best reflect the conditions that you are capturing images in your vineyards.
Getting Started
You’ll need a few things to get started:
Setup a myEV Collector
Data collectors are used in myEV to collect data using mobile devices in the field. As part of this project, we have built a computer vision integration with Roboflow that lets you leverage any open and available model on the Roboflow Universe—a collection of publicly available datasets and models. In this tutorial we will be using ‘nighttime-grape-flower-clusters/1’ which is a nighttime cluster counting model and performed in our trials with the highest levels of accuracy. You can explore our other cluster counting models, leverage any Roboflow Universe models, or even train your own model.
Credit: The data used to train ‘nighttime-grape-flower-clusters/1’ was originally captured and trained by Jonathan Jaramillo as part of his PHD research which motivated the creation of this guide. Read the original paper >>
To setup the data collector:
Login to your myEV farm. We recommend doing configuration on a desktop computer, though you can set this up on your phone as well.
Select the ‘Collect Data’ button. Then click ‘New Data Collector’.
Give your collector a name and then select ‘Add a Header’.
Give your new header a label and select ‘Computer Vision’ under the ‘Header Type’ field. Enter your Roboflow API Key and the Model URL as shown in the images below. Our nighttime cluster counting model’s URL is
https://detect.roboflow.com/nighttime-grape-flower-clusters/1.Once you’ve updated these details, click ‘Add Header’ and then ‘Save’ the data collector. You are now ready to use this data collector to take images and count clusters in the field!
Use your Cluster Counter in the Field
Note: ‘nighttime-grape-flower-clusters/1’ is a nighttime-only model. This means you’ll need to capture vine images at night using a lighting source. We recommend placing light sources a few feet above and/or below your camera as in the figure below:
Here is an example image that we will be using in this tutorial to count clusters:
You now have a handy way to take pictures of vines and automatically count visible clusters. Be sure to install myEV on your phone and head out to the field. Once in front of a vine that you’d like to count clusters on follow these steps:
Open myEV and tap the ‘Collect Data’ icon in the lower-right.
Select your cluster counting collector.
Select ‘Choose file’ under your cluster counter header. This will prompt you to either choose a photo or use your camera to take an image.
The resulting image will upload, process, and resolve with a cluster count and marked annotations of clusters.
Remember to tap ‘Submit’ to finalize the datapoint. You can repeat this process to collect counts across a spatial area. The resulting dataset can be visualized and analyzed within myEV. See our tutorials for more information on mapping in myEV.
Always test and check the accuracy of these tools! We recommend dedicating time to ground-truthing the results of these computer vision tools by manually counting clusters and comparing the model’s inferences vs the actual cluster count. This ratio can be used to make better predictions of clusters in the field.
Next Steps