DOC_05

What's New in This Fork

This page explains all the changes that has been made since forking from the original ChronoRoot repository

Smarter and More Accurate Plant-Recognition Model

Every measurement this tool produces (root length, hypocotyl length, number of side roots) starts from the software correctly telling apart the different parts of the plant in a photo. That’s the job of the segmentation model, and it’s the piece that’s changed the most.

Better segmentations of various phenotypes, especially seedlings with elongated hypocotyls. The original model often mislabeled long hypocotyls , which cut the hypocotyl short and threw off its length measurement. The new model segments these hypocotyls correctly.

Old Model
New Model
A properly weighted moving average for smoothing between frames. After the model labels each photo, the software blends neighboring frames over time so a single bad frame doesn’t make the labels flicker. This blend used to add each new frame on top of a running total, so the total kept growing. It is now a true exponential moving average: each new frame counts for 1 − alpha and the history for alpha. The same alpha setting now behaves the same way for every frame.

More detail on how the model gets retrained is on the training page; the technical class-label reference lives on the architecture notes page.

Native MacOS Support

This project can now be installed and ran natively on MacOS machines

One installer does the whole setup. Running installer_conda_mac.sh installs the Python environment the apps need (including Miniconda, if it isn’t already there), the QR-code library, and the segmentation model.
Real Mac apps to click on. The installer puts ChronoRoot Analysis, ChronoRoot Segmentation and ChronoRoot Image Aligner in your Applications folder, with shortcuts on your Desktop. There’s no terminal needed day to day.
Faster segmentation on Apple Silicon. On newer Macs (M1, M2, M3 and later), segmentation runs on the Mac’s graphics chip instead of only the processor. Older Intel Macs still work, running on the processor.
Updating is one step. Re-running the installer offers to download the latest version first, then brings everything else up to date to match.

Step-by-step instructions are in the install guide.

More Reliable Overall Pipeline

New: Automatic Image Alignment

This project now ships an imageAligner app that corrects potential camera drift automatically before any measurement happens, using small printed markers placed in view of the camera as reference points.

Fewer Silent Mistakes in the Measurements

A number of bugs that produced wrong numbers, with no error message at all, have been found and fixed:

A neighboring plant’s hypocotyl could get measured by mistake. If a plant’s own hypocotyl wasn’t visible yet, the software could grab a nearby plant’s instead, since it just picked the biggest matching blob it could see. It now anchors on where that specific plant’s seed was originally placed, so it can no longer borrow a neighbor’s growth.

Before the fix: the hypocotyl measurement (yellow) is drawn on the neighboring plant to the right, while the tracked plant's own root is shown in red — Before

After the fix: the hypocotyl measurement (yellow) sits on the tracked plant itself, directly above its red root — After

Hypocotyl length no longer gets wiped to zero early in an experiment. Hypocotyl growth is measured independently of the root, but a bug caused it to be reported as 0 on any frame where the root hadn’t been successfully detected yet, even though the real hypocotyl measurement was sitting right there. That data is no longer thrown away.
Improved root tracking algorithm The software used to treat the lowest point in the image as the tip of the main root. A side root that dipped below it, or a main root that curved sideways near the end, could make the software follow the wrong root and report its length as the main root length. The main root is now measured as the longest path from the root origin to a root tip, so it no longer depends on which root happens to sit lowest.
Growth charts no longer come out empty or cut short for experiments that don’t photograph every 15 minutes. The chart-building step assumed a fixed 15-minute interval between photos; any other interval (say, every 3 hours) used to produce truncated or blank charts. It now adapts to whatever interval was actually used.
Growth curves no longer flatten out artificially at the end. A smoothing step used to fabricate zero-valued data past the edge of the real measurements, which dragged down and flattened the last few points of every curve. The smoothing now stops at the real edge of the data instead of padding past it.
.tif/.tiff photos are recognized. Experiments captured in those formats used to be silently skipped entirely, because the software only looked for .png files.

Known Issues Still Being Worked On

Nothing hides bugs here. These are known and tracked openly:

Very small, newly-emerged roots or hypocotyls can occasionally disappear from a few frames of measurement before they’ve grown large enough to be reliably tracked. It usually self-corrects as the structure gets bigger.
A step meant to clean up false-positive short roots detected right at the start of an experiment can, in rare cases, mistakenly wipe out real earlier measurements later in the same timelapse too. This is suspected but not yet confirmed as a cause of some “root length not calculated” reports.