Stage 4: Rendering

Synthesize novel-view images from the trained 3DGS model.

---

What This Stage Does

graph LR
    A[✨ point_cloud.ply<br/>Trained model] -->|render.py| B[📷 Train Views<br/>train/ folder]
    A -->|render.py| C[📷 Test Views<br/>test/ folder]
    B --> D[📊 PSNR · SSIM · LPIPS]
    C --> D
    style A fill:#e1f5ff
    style B fill:#e1ffe1
    style C fill:#e1ffe1
    style D fill:#fff3e0

Estimated time: ~5 minutes

---

Command

conda activate 3dgs

python render.py \
    -m /path/to/date_20260119/output \
    --iteration 30000

Argument	Description
-m	Path to your trained model folder
--iteration	Which checkpoint to render from (use 30000 for final)

---

Monitoring Render Progress

📸 Screenshot to capture

Screenshot the terminal while render.py runs — it shows each camera view being rendered with a progress counter.

Rendering terminal — typically renders 329 train views + test views in ~5 minutes

Expected terminal output:

Loading trained model at iteration 30000
Rendering train views: 100%|████████| 263/263 [03:12<00:00]
Rendering test views:  100%|████████| 66/66  [00:48<00:00]

---

Output Structure

ls output/train/ours_30000/
# gt/        ← ground truth images (original frames)
# renders/   ← model-rendered images

ls output/test/ours_30000/
# gt/
# renders/

📸 Screenshot to capture

Screenshot the output directory tree and also open one rendered image side-by-side with its ground truth.

Output folder structure after rendering — both gt/ and renders/ should have equal image counts

---

Visual Quality Check

Ground truth (left) vs 3DGS rendered (right) — scrolling through 1,643 training views. At PSNR 23.71 dB, differences are barely visible.

Static comparison: ground truth frame vs 3DGS render at PSNR 23.71 dB

Good Render ✅Poor Render ❌

• Sharp leaf edges
• Accurate color reproduction
• Stem structure clearly defined
• Minor noise in background only

• Blurry or smeared leaves
• Floaters (spurious splats in mid-air)
• Missing plant regions
• Incorrect colors

---

Quality Metrics

After rendering, evaluate metrics:

python metrics.py -m /path/to/date_20260119/output

📸 Screenshot to capture

Screenshot the metrics.py output showing PSNR, SSIM, and LPIPS values.

Metrics output — compare against our validated benchmarks in the table below

Metric	Our Result	Minimum Acceptable
PSNR	23.71 dB	> 20 dB
SSIM	0.82	> 0.75
LPIPS	0.18	< 0.30

What these metrics mean

• PSNR (Peak Signal-to-Noise Ratio): Higher is better. > 23 dB is excellent for plant scenes.
• SSIM (Structural Similarity): 0–1, higher is better. Measures structural fidelity.
• LPIPS (Perceptual Similarity): Lower is better. Measures perceptual difference.

---

How Renders Enable Trait Extraction

The key insight: rendered images are in normalized image space — the plant always occupies a consistent portion of the frame regardless of capture date. This is what enables scale-invariant trait extraction.

Scale inconsistency in PLY coordinates (left) vs consistency in rendered image space (right) — this is why rendering is critical before trait extraction

Why PLY Heights Are Meaningless Without Calibration

COLMAP reconstructions are up-to-scale: each date gets its own independently-scaled coordinate system. The raw PLY height is in arbitrary world units — not metres. Multiplying by a per-date scale factor still produces wildly inconsistent results, making temporal comparison impossible directly from point clouds.

The rendered image height (fraction of frame) is scale-invariant and consistent across all dates.

The core problem this pipeline solves

Using PLY coordinates directly for plant height measurement gives nonsensical results across dates — heights of 0.81 m, 8.24 m, and 11.17 m for the same plant growing smoothly over a few days. Rendering to a fixed viewpoint eliminates this entirely.

Date	Raw PLY height	COLMAP scale	"Real" (unreliable ❌)	Rendered height ✅
Jan 19	9.12 (arb.)	0.089	~0.81 m	85.5% of frame
Jan 21	16.35 (arb.)	0.504	~8.24 m ❌	89.3% of frame
Jan 23	12.82 (arb.)	0.486	~6.23 m ❌	70.3% of frame
Jan 26	19.82 (arb.)	0.563	~11.17 m ❌	89.3% of frame
Jan 28	7.42 (arb.)	0.396	~2.94 m ❌	84.7% of frame
Jan 30	6.53 (arb.)	0.495	~3.24 m ❌	89.8% of frame
Feb 2	8.87 (arb.)	0.567	~5.03 m ❌	62.8% of frame
Feb 4	9.12 (arb.)	0.544	~4.96 m ❌	74.4% of frame
Feb 6	7.94 (arb.)	0.539	~4.28 m ❌	89.9% of frame
Feb 9	8.95 (arb.)	0.450	~4.03 m ❌	89.7% of frame
Feb 11	7.08 (arb.)	0.461	~3.26 m ❌	88.3% of frame
Feb 13	5.06 (arb.)	0.538	~2.72 m ❌	85.9% of frame
Feb 16	10.44 (arb.)	0.484	~5.06 m ❌	86.1% of frame
Feb 18	6.81 (arb.)	0.466	~3.17 m ❌	82.4% of frame
Feb 20	6.05 (arb.)	0.483	~2.93 m ❌	67.6% of frame
Feb 23	4.74 (arb.)	0.577	~2.74 m ❌	89.9% of frame
Feb 25	16.57 (arb.)	0.462	~7.66 m ❌	90.0% of frame
Feb 27	20.81 (arb.)	0.524	~10.91 m ❌	77.9% of frame
Mar 2	9.12 (arb.)	0.543	~4.95 m ❌	89.5% of frame
Mar 4	10.53 (arb.)	0.511	~5.38 m ❌	88.4% of frame
Mar 6	8.88 (arb.)	0.549	~4.87 m ❌	89.1% of frame
Mar 9	10.32 (arb.)	0.506	~5.22 m ❌	89.9% of frame

Lesson: The rendered height_norm values are stable (62–90% of frame, reflecting real plant growth). The PLY-derived heights are chaotic and unusable for temporal analysis without a scene-level metric calibration reference.

---

Batch Rendering (Multiple Dates)

#!/bin/bash
# batch_render.sh

for MODEL_DIR in data/*/output; do
    DATE=$(basename "$(dirname "$MODEL_DIR")")
    echo "Rendering $DATE..."
    python render.py -m "$MODEL_DIR" --iteration 30000
    echo "✅ $DATE rendering complete"
done

---

Next Step

With renders in output/train/ours_30000/renders/, proceed to trait extraction.

→ Stage 5: Trait Extraction