Applied computer vision / Master’s thesis / 2026

TomatoVision

One scene.
Different hypotheses.

A computer-vision study that detects three greenhouse tomato maturity stages and compares a YOLOv11 baseline with modified models and a three-model WBF ensemble.

YOLOv11 baseline detections on a photograph of a tomato plant in a garden greenhouse — **YOLOv11 baseline**Qualitative demo · not validation imagery

The red fruit is classified as orange in this scene. Domain shift is shown as observed, not retouched. Photo: Kolforn; detections overlaid. Image adaptations shared under CC BY-SA 4.0.

My contribution: Computer vision research · model development · evaluation
Built with: YOLOv11 · Swin Transformer · PyTorch · OpenCV · Weighted Boxes Fusion

The problem

Start with the work.

Greenhouse tomatoes overlap, hide behind leaves, and appear at different scales, making maturity detection difficult for a single detector configuration.

What I built

Adjie evaluated a YOLOv11 baseline, Swin Transformer and multi-scale SPPF variants, then fused their predictions with Weighted Boxes Fusion while reporting single-model and ensemble results separately.

01 / Research method

Change the architecture.
Then test the ensemble.

Baseline
YOLOv11
A fixed reference for the architectural comparisons.
Single-model variants
Swin-T
+ multi-scale SPPF
Evaluate the modified detector independently.
Prediction fusion
Weighted
Boxes Fusion
Combine model outputs and report the ensemble separately.

02 / Quantitative evaluation

The result depends
on what you run.

Single-model and ensemble results from the recorded thesis evaluation. These figures do not measure the demo photograph above.

Seven evaluated configurations
Configuration	Type	mAP@0.5	mAP@0.5:0.95	ms / image
YOLOv11	single	0.795	0.470	55.28
+ Swin-T	single	0.805	0.474	53.46
+ Swin-T + MS-SPPF	single	0.807	0.477	52.77
Combine 1YOLOv11 + Swin-T	ensemble	0.814	0.490	69.88
Combine 2YOLOv11 + Swin-T+MS-SPPF	ensemble	0.817	0.492	69.37
Combine 3Swin-T + Swin-T+MS-SPPF	ensemble	0.812	0.487	67.17
Combine 4all three	ensemble	0.824	0.499	88.93

Timing belongs to the recorded evaluation setup; it is not a browser benchmark or a deployment guarantee.

03 / Dataset & training

Keep the denominator
in the story.

Source imagery: 1,051 images · Known-You Seed Co., Pingtung, Taiwan
Source annotations: 12,168 green · 14,640 orange · 13,989 red
Exported split: exported 2,193 train (augmented) / 160 val / 160 test at 1280×1280
Training configuration: 1088px · batch 4 · 400 epochs · patience 100

Evidence / verification

The public record.

YOLOv11 baseline: 0.795 mAP@0.5
Best modified model: 0.807 mAP@0.5
Three-model WBF: 0.824 mAP@0.5
WBF stricter metric: 0.499 mAP@0.5:0.95

Start with the work.

YOLOv11

Swin-T+ multi-scale SPPF

WeightedBoxes Fusion

Keep the denominatorin the story.

Swin-T
+ multi-scale SPPF

Weighted
Boxes Fusion

Keep the denominator
in the story.