Pine Tree Detection with YOLOv5

A lightweight pipeline for detecting pine trees in RGB images using YOLOv5. This repository covers:

• Converting Mask R-CNN JSON annotations to YOLO format
• Organizing a custom dataset for YOLOv5
• Running a series of training experiments
• Comparing results (mAP, recall, precision, confusion matrix)
• Analysis of performance and next steps

📂 Repository Structure

. ├── data │ ├── train │ │ ├── images │ │ └── labels │ └── validations │ ├── images │ └── labels ├── yolov5 │ ├── train.py │ ├── val.py │ ├── detect.py │ ├── dataset.yml │ └── … (YOLOv5 code & configs) ├── convert_to_yolo.py # Script to turn Mask R-CNN JSON → YOLO TXT └── README.md

🗂 Dataset

• One class: pine_tree (class 0).

• Structure (root: data/):

data/ ├─ train/images ├─ train/labels ← YOLO TXT files ├─ validations/images └─ validations/labels

• yolov5/dataset.yml:

yaml # root: yolov5/ path: ../data train: train/images val: validations/images nc: 1 names: 0: pine_tree

🔄 Annotation Conversion

Use convert_to_yolo.py to turn your Mask R-CNN JSONs into YOLO .txt label files:

```bash python converttoyolo.py \ --jsondir data/train/annots \ --labelsdir data/train/labels

python converttoyolo.py \ --jsondir data/validations/annots \ --labelsdir data/validations/labels ```

⚙️ Training Experiments

We ran six distinct experiments, each saved under runs/train/<name>/:

| Name | Model | Key Trick | | --------------- | -------- | ---------------------------------------- | | pine_tree_run | YOLOv5-s | default training (50 epochs) | | exp_freeze | YOLOv5-s | first 30 ep freezing backbone → +20 ep | | exp_unfreeze | YOLOv5-s | fine-tune all layers | | exp_anchor | YOLOv5-s | (auto)recomputed anchors | | exp_evolve | YOLOv5-s | hyperparameter evolution (--evolve) | | exp_yolov5m2 | YOLOv5-m | medium model + best exp_evolve hparams |

Each was launched with a one-liner, e.g.:

bash python yolov5/train.py \ --weights yolov5s.pt \ --data dataset.yml \ --img 640 \ --batch 16 \ --epochs 50 \ --name exp_anchor

📊 Results

Quantitative Metrics

| Experiment | Precision | Recall | mAP\@50 | mAP\@50-95 | “Accuracy” (TP/(TP+FP+FN)) | | --------------- | :-------: | :----: | :-----: | :--------: | :------------------------: | | pine_tree_run | 0.664 | 0.600 | 0.646 | 0.417 | 46.2 % | | exp_yolov5m2 | — | 0.830 | — | — | 41.5 % |

(Detailed logs are in each run’s results.txt.)

Best Confusion Matrix

• True Positive Rate (recall): 83 %
• False Negative Rate: 17 %
• False Positive Rate (background→pine): 100 %
• Implicit background class yields a 2×2 matrix for single‐class detection.

🖼️ Example Detections

Here are two sample detection outputs:

🔍 Analysis

1. Dataset Challenges

• Size & diversity: only ~150 train images with limited backgrounds
• Image quality: some images are low-contrast or blurred
• Single‐class, implicit negatives: no explicit “background” boxes

1. Model Capacity & Augmentation

• Upgrading to YOLOv5-m and applying evolved hyperparameters boosted recall from 60 % → 83 %.
• Precision remains limited by high false positives on complex backgrounds.

1. Detection‐Accuracy vs. Classification

• In object detection, we measure precision/recall/mAP; the “accuracy” here is TP/(TP+FP+FN) ≈ 41 %.

🚀 Next Steps

• Expand & diversify data: collect more scenes (different lighting, angles, seasons).
• Improve annotations: tighter, more consistent bounding boxes; consider adding a few explicit “background” crops for negative sampling.
• Multi-spectral inputs: include NIR or depth channels if available.
• Advanced augmentation: copy-paste, stronger color/scale/perspective variants.
• Ensemble & TTA: combine multiple runs or test-time augmentations for gain.

▶️ Usage

1. Install dependencies

bash pip install -r yolov5/requirements.txt 2. Convert annotations

bash python convert_to_yolo.py ... 3. Train

bash python yolov5/train.py --data yolov5/dataset.yml --name your_experiment 4. Validate

bash python yolov5/val.py --weights runs/train/your_experiment/weights/best.pt \ --data yolov5/dataset.yml \ --save-conf --verbose --name val_your_experiment 5. Detect

bash python yolov5/detect.py --weights runs/train/your_experiment/weights/best.pt \ --source path/to/images \ --save-txt --name detect_your_experiment

Acknowledgments

• Ultralytics YOLOv5
• Early experiments with Mask R-CNN annotations