log in | about 

Believe it or not, the USA is the largest consumer of wine that guzzles more than 10% of all wine produced on the planet. However, it lags somewhat in production. Turns out that maximizing grape yield relies heavily on measurements during the growing season, in particular, on crop estimation. If certain areas are underperforming, it is often possible to fix the issue by, e.g., additional irrigation and fertilization.

Crop estimation is an expensive labor-intensive process that was previously carried out only by humans. The Robotics Institute of Carnegie Mellon University (in collaboration with Cornell University and stakeholders) works on developing automated measuring techniques. At Carnegie Mellon University, the group is lead by Stephen Nuske.

What is truly astonishing is that the proposed technology relies only on passive vision techniques, which are considered unreliable to be used outside a lab. Unlike self-driving cars requiring expensive laser-powered sensing devices called LIDARs, the proposed technology uses only a camera. The camera resides on a small cart that drives at a speed of about 5mph (if I remember correctly, there is also a flash to neutralize variability in lighting). While driving, the camera makes overlapping pictures of grape vines. Obtained images are processed to detect individual grapes and count them!

Although image recognition algorithms have reached a certain level of maturity, it is still challenging to detect individual grapes, because there are millions of potential locations to check in a single picture. This is especially hard when grapes did not ripen (and consequently both leaves and grapes are green). However, the researchers from the Robotics Institute of Carnegie Mellon University can count grapes even in real time! To accomplish this complex task, they use a combination of a quick high-recall low-precision filtering algorithm and a more accurate algorithm that removes false matches. The high-recall low-precision algorithm is an ensemble of two relatively simple key-point detection algorithms. The approach is described in a series of publications. The overall accuracy seems to be pretty good and the technology might be commercialized in not-so-distant future.

To conclude, I would like to note that, in addition to grape counting in your vineyard, Stephen Nuske worked on several other cool projects, where passive vision was applied to real-world problems. These may be interesting to both practitioners and lab scientists specializing in computer vision.