Skip to content

Results & Validation

Research results and validation. All values match the submitted manuscript (Computers and Electronics in Agriculture).

Reconstruction Quality Metrics (Condition 1, 5 fps, Jan 19)

Metric Value Interpretation
PSNR 23.71 dB Good (>20 dB acceptable)
SSIM 0.773 Good (1.0 = perfect)
LPIPS 0.222 Good (0.0 = perfect)
PSNR temporal CV 3.5% Excellent consistency across 22 sessions
Mean PSNR (22 sessions) 23.84 dB All sessions exceed 22.5 dB

Plant Trait Monitoring Results (cross-session consistency)

Why CV is the right metric

This is a monitoring system: the goal is to track relative change over time, not to report absolute, calibrated trait values. The coefficient of variation (CV, lower = more consistent) is therefore the primary and appropriate evaluation metric.

Metric Proposed (render-space) Baseline (direct PLY) Improvement
h_norm CV 9.8% 28.0% 2.86× more stable
a_norm CV 16.7% 22.6% (PLY spread) More stable
Scale-calibrated PLY CV 35.1% Worse than direct PLY

Biological Validation — Pruning-Event Detection

The render-space signal detects real crop-management events, which is the key evidence that it tracks biology rather than noise.

Aspect Result
Documented pruning events detected 3 / 3
Mean drop at pruning ≈ 0.20 normalized height (15–27 pp)
Significance (pruning vs. non-pruning dates) p = 0.0008
Background variation 9.8% (CV) — drops are well above it

Physical-Reference Validation — 45 cm Pipe

Render-space height is corroborated against an in-scene 45 cm bench pipe used as a ruler, h_gt = (plant_px / pipe_px) × 45 cm, annotated across all 22 sessions.

Aspect Result
Correlation (render-space vs. pipe reference) r = 0.74, p < 0.001
0.55
Reference mean height 159.9 cm
Spread h_gt CV 8.9% vs. h_norm CV 9.6%

Corroboration, not accuracy

r = 0.74 shows the signal is grounded in real-world scale — it is not a claim of absolute-height accuracy. The ~45% unexplained variance is noise in the pipe reference itself (SfM scale ambiguity across sessions), not pipeline error.

Plant Growth Time-Lapse

3DGS-rendered plant growth across all 22 acquisition dates (Jan 19 – Mar 9, 2026). Each frame shows the height measured from the rendered canonical image.


Validation Results

  • Dates: 22 capture dates
  • Duration: 49 days (Jan 19 – Mar 9, 2026)
  • Success Rate: 100% (22/22)
  • Computing: DNN19, RTX 6000 Ada

Key Finding

Fixed viewpoint (Condition 1, 5 fps) outperforms all other acquisition conditions:

Condition PSNR
Cond. 1 — Fixed end-a, lower viewpoint 23.71 dB
Cond. 5 — One-way a→d, tier 1 21.42 dB
Cond. 2 — Fixed end-b, lower viewpoint 20.11 dB
Cond. 3 — Round-trip, tier 1 19.37 dB
Merged (all 6 combined) 18.60 dB

Lesson: Acquisition consistency > method sophistication


Limitations and Future Ground Truth

Honest limitations

  • No calibrated ground truth (RMSE): these 22 sessions have no tape-measured physical heights, so the system is validated for consistency (CV) and change detection, not absolute accuracy. The 45 cm-pipe result is corroboration, not calibration.
  • Reference-instrument noise: the pipe reference carries its own SfM scale ambiguity (pipe spans 263–420 px for a fixed 45 cm), which accounts for most of the unexplained variance in the r = 0.74 agreement.
  • Operator not blinded to the pruning schedule; events were confirmed from work logs post-hoc.

Planned ground-truth protocol (future work)

A three-tier protocol would upgrade the evaluation from consistency to absolute accuracy:

  1. Tape measurements on 5–10 tagged plants per session (direct RMSE).
  2. Rigid calibration bar of known length in-scene — removes the dominant variance source of a movable pipe.
  3. ArUco markers for a full per-session similarity transform → absolute metric units.