NPEC — research (in progress) Publication

4-Channel RT-DETR for Overlapping Roots

Research in progress

Ongoing research tackling a known blind spot in high-throughput phenotyping: current frameworks discard any image with touching or overlapping plants. I'm testing whether encoding root-intersection points as a spatial topological prior in a 4-channel RT-DETR can recover that lost data.

Client

NPEC — research (in progress)

Year

2026

Domain

Vision, Research

Stack

PyTorch · RT-DETR · SegFormer · OpenCV

Context

This is the next phase of my root phenotyping work for NPEC, and it is actively in progress. The earlier system assumed roots don’t overlap — a fair assumption early in an experiment, but one that breaks down fast.

The gap

State-of-the-art phenotyping frameworks are excellent at automating analysis, but in high-throughput mode they discard every image containing at least one touching or overlapping plant. For Arabidopsis thaliana, plants begin overlapping around day 30 — and experiments often run 50–60 days — so a large, valuable slice of the data is simply thrown away.

The core difficulty is optical: at a root-crossing point, the RGB values and root widths are almost identical to ordinary, un-entangled root, so a standard detector has nothing obvious to separate one plant from another.

Research question

How does integrating a spatial topological layer — encoding root-intersection points as geometric priors — affect the mean Average Precision (mAP) of an RT-DETR model when identifying individual plants in high-density Petri dishes with moderate-to-heavy root crossing?

Status

Currently in the research phase: preliminary exploratory analysis is done, the gap is characterised against existing frameworks, and the experimental design around the 4-channel RT-DETR approach is underway. Results and a write-up will follow — this card will be updated as the work progresses.