Custom pipelines · Methods & evidence
Reading this
potato time series.
24 fixed-pocket plants. Three capture dates. Source-linked measurements and provisional visible blade estimates.
The collection
Six blue four-pocket trays were scanned on 7, 9 and 13 October 2026: 18 scans and 72 plant observations. The experimenter confirmed that plants were not rearranged between pockets. Days 0, 2 and 6 are measured from the first scan, not from the start of a treatment.
Trays 2169, 2441 and 2626 belong to the high-light chamber. Trays 2625, 2629 and 2630 belong to the low-light chamber. Tray 2636 is unrelated and excluded.
Position names refer to an observer looking into the tray from its label end. Plants retain their pocket identity across dates. Individual leaf identities are not tracked.
What the measurements mean
| Height · cm | H99.5: the 99.5th percentile of observed source-centre height above each pocket’s local substrate, after normalization into 0.5 mm voxels. This is not the translucent outer edge of a splat. |
|---|---|
| Apparent canopy footprint · cm² | Orthographic Gaussian canopy area at accumulated alpha ≥ 0.5, on a 0.5 mm grid. It describes an apparent footprint, not total anatomical leaf area. Alpha-boundary and sampling sensitivity remain part of its interpretation. |
| Convex envelope · cm³ | The convex boundary around the retained plant centres. It includes empty space between leaves. It is not tissue volume or biomass. |
| Visible blade candidates | A provisional SAM2.1 estimate of separately visible leaf blades or surface instances. It is not a validated count of botanically defined whole leaves. |
All chamber summaries on this website are means per plant over 12 plants in three trays. The original dossier also contains mean tray-sum summaries; those are different aggregations.
Scale and reconstruction
Metric scale uses a fixed reference for each tray, inherited from the audited baseline blue-plastic photogrammetry reference. The same reference is carried through the series. Camera-only metric fits passed for 14 of the 18 scans; four were rejected. Camera fitting is not the final scale authority in this analysis.
The blue tray reference, local substrate and label-side orientation connect observed splat centres to the metric frame. This provides study-specific scale grounding; it does not independently establish absolute biological measurement accuracy.
Native Gaussian subsets preserve trainer coordinates. The separate metric point clouds are in metres. The films and common-view renders use the recorded calibration to display the observations at a consistent physical scale.
Visible blades and source images
SAM2.1 Hiera Small was applied to three source-image crops per plant observation: 216 crop inferences in the final CPU run. The reported estimate is the maximum supported single-view count across those views. Counts from different views are never added together. Hidden blades are not inferred.
Of 72 observation records, 67 have at least one model or association review flag. All observations have an estimate; 71 have nonempty candidate sets in all three views. View choice, occlusion and over- or under-segmentation affect the result. Model confidence scores are not measured accuracy.
The explorer exposes each crop and its count overlay, plus the view-specific counts and review flags. Treat coloured regions and numbers as model proposals that can be inspected, not validated anatomical labels.
What the light comparison supports
There was one chamber per light regime: two chambers in total. Plants and trays are subsamples; dates are repeated observations. Differences describe these two chambers and cannot isolate a causal light effect from other chamber differences.
The low-light chamber plants were already taller at baseline. Both chambers increased their apparent canopy footprint during the six-day observation window. Calibrated light intensities, treatment onset and randomization details were unavailable. No treatment p-values or causal confidence intervals are presented.
How to explore the 3D collection
The stage renders the actual recorded Gaussian subsets directly in your browser. Change date to compare three discrete observed states; drag to turn the plants, choose a tray, or select an individual. No intermediate growth geometry or hidden leaves are generated.
Plant anchors and physical scale remain fixed across dates. The collection is arranged for comparison rather than reproducing the chamber layout. The source photograph shows the growth environment.
The web format retains source DC colour, opacity and Gaussian shape, with float32 metric positions and scales. Rotation and opacity are quantized for the standard 32-byte splat format. The maximum observed relative covariance difference from rotation packing is 4.20%; original native PLY sources remain unchanged. Colour is not full view-dependent radiance.
Data and provenance
This display was assembled from the complete Dario / Hester research package dated 25 October 2026. The 72-row joined table is the measurement authority. The website includes all 72 interactive Gaussian subsets and all 24 plant histories, along with apparent canopy masks, three RGB crops and three SAM overlays for each observation, source scene photographs and common-view renders.
The original scan archive, native models, complete 359-page technical dossier, historical diagnostics and processing intermediates remain in the source collection. Web images are resized derivatives with metadata removed. The interactive Gaussian data are calibrated derivatives of the native source subsets, verified against the supplied metric point clouds. The exported measurements retain the source values.
From capture to a new research question
The value of a retained capture is that its source observations can be revisited. Here, photogrammetry supplies the reference, splats supply the recorded visual structure, and a segmentation foundation model adds inspectable blade proposals. Trait definitions and validation remain specific to the research question.
Discuss a custom Pro workflow ↗