Here's the answer - what was the question?

High throughput phenotyping systems with batteries of imaging sensors are becoming commonplace. While of immense potential use in automating phenotyping tasks, and permitting high numbers of experimental replicates at minimal extra cost, they present new challenges reminiscent of those facing “Big Data”. The Lemnatec installation at the UK National Plant Phenomics Centre [NPPC] generates on average 10K images/day, consuming 6-7Gb of store.

This immense resource is most unlikely to be inspected by human observers, and so the traditional approach of making manual observations for phenotyping will not be replicated: rather, we will need automatic image inspection and analysis to provide comparable observations.

Computer vision has for decades considered such problems and been of great service in application disciplines such as medicine and industrial inspection. Frequently however, the computer scientist delivers an observation that the human observer might find difficult to make, while the manual observation might prove hard to automate. Thus the question answered might not be the one asked, and the challenge becomes to demonstrate or verify that the answers are of comparable scientific value.

We will present two case studies illustrating this issue that are derived from the work of the computer vision team at NPPC, trying to answer questions posed by plant biologists. We conclude that there is much the two disciplines can say to and do for each other.