BloomNet: Exploring Single vs. Multiple Object Annotation for Flower Recognition Using YOLO Variants

This paper introduces the FloralSix dataset and benchmarks various YOLO architectures to demonstrate that annotation density and model selection significantly impact flower recognition performance, with precision-optimized models excelling in sparse scenarios and recall-optimized models outperforming in dense, multi-object environments.

The Gist

Imagine you are a botanist trying to count flowers in a massive, wild garden. Sometimes, the flowers are spread out like stars in a clear night sky, easy to spot one by one. Other times, they are packed so tightly together in a bush that they overlap, hide behind each other, and form a colorful, chaotic mess.

This paper, titled "BloomNet," is like a report card for a team of high-tech "flower counters" (AI cameras) trying to solve this exact problem. The researchers wanted to know: Does it matter how we teach the AI to look at the flowers?

Here is the story of their experiment, broken down into simple concepts:

1. The New "Flower Library" (The Dataset)

Before the AI could learn, the researchers needed a library of pictures. They created a new collection called FloralSix, which contains nearly 3,000 high-definition photos of six different types of flowers (like Hibiscus and Marigolds) from gardens in Bangladesh.

They did something clever with these photos: they labeled them in two different ways, like teaching a student two different study methods:

• Method A (The "Spotlight" Approach): They drew just one box around the most obvious flower in the picture. This is like telling the AI, "Ignore the crowd, just find the star of the show."
• Method B (The "Crowd Control" Approach): They drew many boxes, labeling every single flower visible, even the tiny ones hiding in the back or overlapping with others. This is like telling the AI, "Count everyone in the room, no matter how crowded it gets."

2. The Contestants (The AI Models)

The researchers pitted several versions of a famous AI eye-spy called YOLO (You Only Look Once) against each other. Think of these models as different types of detectives:

• YOLOv5s: The veteran detective. Fast and reliable, but maybe a bit old-school.
• YOLOv8 (n, s, m): The new generation. They come in different sizes: "Nano" (tiny and fast), "Small" (balanced), and "Medium" (stronger but heavier).
• YOLOv12: The futuristic super-detective, designed specifically to handle tiny objects and crowded scenes.

3. The Experiment: How They Learned

The AI models were trained using a "coach" called an optimizer. The researchers tested two coaches:

• SGD: The steady, disciplined coach who takes small, consistent steps.
• AdamW: The flashy, high-speed coach who tries to sprint but sometimes stumbles.

The Result? The steady coach (SGD) won every time. It helped the AI learn more reliably, just like a student who studies a little bit every day usually does better than one who crams all night.

4. The Big Findings

Here is what happened when the AI took the test:

• Scenario 1: The "Spotlight" Test (Sparse Flowers)
  When flowers were spread out, the YOLOv8m (Medium) model was the champion. It was incredibly precise, almost never missing a flower or confusing one for another. It was like a sniper: slow to aim, but perfect when it fired.

  • Takeaway: If you just need to find one main flower in a field, use the Medium model with the steady coach.

• Scenario 2: The "Crowd" Test (Dense Flowers)
  When the flowers were packed tight like sardines, the YOLOv12n (Nano) model surprised everyone. Even though it was the smallest model, it was the best at finding every flower in the mess. It didn't get confused by the overlapping petals.

  • Takeaway: If you are looking at a dense bush and need to count every bloom, the tiny, futuristic model is your best friend.

5. Why Does This Matter? (The Real-World Magic)

Why do we care about counting flowers? Because this technology is the brain behind smart farming.

• Robotic Pollinators: Imagine a robot bee that needs to know exactly where a flower is to pollinate it. If the flowers are crowded, the robot needs the "Crowd Control" AI to avoid crashing.
• Yield Prediction: Farmers need to know how many flowers are on a plant to guess how much fruit or seed they will harvest.
• Stress Detection: If a plant stops flowering, it might be sick or thirsty. The AI can spot this early.

The Bottom Line

This paper teaches us that one size does not fit all.

• If you want to find a single, clear flower, use a medium-sized AI trained to look for one target.
• If you are staring into a dense, messy bush, use a tiny, super-smart AI trained to see the whole crowd.

It's like choosing the right tool for the job: you wouldn't use a sledgehammer to crack a nut, and you wouldn't use a tiny screwdriver to build a house. The researchers found the perfect tool for every type of flower garden.

Technical Summary

1. Problem Statement

The paper addresses the critical need for automated flower recognition in precision agriculture, specifically for tasks like plant phenotyping, crop estimation, and yield monitoring. Existing challenges include:

• Complexity of Natural Scenes: Flowers vary in orientation, illumination, and background clutter.
• Object Density: Traditional single-label classification fails when multiple flowers appear in a single frame, especially when they are overlapping, partially hidden, or forming dense clusters.
• Dataset Limitations: Existing datasets (e.g., Oxford102, Oxford17) focus primarily on classification with limited or inconsistent bounding box annotations, making them unsuitable for analyzing object density or robust multi-flower detection.
• Annotation Strategy Gap: There is a lack of research comparing how different annotation densities (single vs. multiple bounding boxes per image) impact the performance of state-of-the-art object detection models.

2. Methodology

Dataset: FloralSix

The authors introduced a new dataset named FloralSix, comprising 2,816 high-resolution images (3000x4000 pixels) of six flower species collected from gardens in Bangladesh: Hibiscus, Marigold, Zinnia, Melampodium, Crape Jasmine, and Madagascar Periwinkle.

• Annotation Regimes: The dataset was annotated using the Roboflow tool under two distinct strategies to test model sensitivity to object density:
  1. SISBB (Single Image Single Bounding Box): Each image contains one bounding box highlighting the main flower. This serves as a baseline for sparse/isolated scenarios.
  2. SIMBB (Single Image Multiple Bounding Box): Every visible flower in an image is annotated. This creates a dense dataset with 6,934 total bounding boxes (averaging 2–26 boxes per image), simulating crowded, real-world agricultural scenes.
• Split: The data was divided into 75% training, 15% validation, and 15% testing.

Model Selection

The study benchmarked five YOLO (You Only Look Once) variants known for real-time performance:

• YOLOv5s: A lightweight baseline.
• YOLOv8n, s, m: Variants with improved CSP backbones and necks for better feature reuse.
• YOLOv12n: An anchor-free architecture designed specifically for small objects and dense environments.

Training Configuration

• Preprocessing: Images were resized to 640x640 pixels.
• Optimizers: Models were trained using both Stochastic Gradient Descent (SGD) and AdamW to determine the optimal optimizer.
• Hyperparameters: Initial learning rate of 0.01, batch size of 16, and 100 epochs with early stopping.
• Loss Functions: A combination of classification loss, objectness loss, and bounding box loss (using CIoU).

Evaluation Metrics

Performance was measured using standard object detection metrics:

• Precision & Recall
• Mean Average Precision (mAP) at IoU thresholds of 0.5 (mAP@0.5) and 0.5–0.95 (mAP@0.5:0.95).

3. Key Contributions

1. FloralSix Dataset: The creation and public release of a fully labeled, high-resolution flower dataset specifically designed for object detection, covering six species with both sparse and dense annotation scenarios.
2. Annotation Density Analysis: A novel comparative study analyzing how SISBB vs. SIMBB annotation strategies affect model performance, revealing that recall-optimized models excel in dense environments while precision-oriented models suit sparse scenarios.
3. Optimizer Benchmarking: A comprehensive evaluation showing that SGD consistently outperforms AdamW across all YOLO variants and annotation types for flower detection.
4. Model Benchmarking: Establishing the first experimental baseline for flower object detection using YOLOv5, v8, and the emerging YOLOv12 architectures.

4. Results

Performance under SISBB (Sparse/Isolated)

• Top Performer: YOLOv8m (SGD) achieved the best results.
  • Precision: 0.956
  • Recall: 0.951
  • mAP@0.5: 0.978
  • mAP@0.5:0.95: 0.865
• Observation: In sparse scenarios, larger models (v8m) with SGD optimization provided the highest accuracy in isolating single objects.

Performance under SIMBB (Dense/Clustered)

• Top Performer: YOLOv12n (SGD) outperformed other models in the complex, multi-object scenario.
  • Precision: 0.924
  • Recall: 0.869
  • mAP@0.5: 0.934
  • mAP@0.5:0.95: 0.752
• Observation: While overall metrics dropped compared to SISBB due to occlusion and density, YOLOv12n demonstrated superior robustness in detecting multiple overlapping instances.

Key Findings

• Optimizer Impact: SGD consistently yielded better convergence and higher metrics than AdamW across all models and scenarios.
• Density Sensitivity:
  • Sparse Environments: Benefit from precision-oriented, slightly larger models (e.g., YOLOv8m).
  • Dense Environments: Benefit from recall-optimized, anchor-free architectures (e.g., YOLOv12n) capable of handling occlusion.
• Trade-offs: YOLOv8m (49.6 MB) is ideal for high-accuracy single-object tasks, while the smaller YOLOv12n (5.6 MB) offers a better balance of speed and accuracy for dense, multi-object detection in resource-constrained environments.

5. Significance and Applications

This research bridges a significant gap in agricultural computer vision by providing a robust framework for non-destructive crop analysis. The findings have direct implications for:

• Robotic Pollination: Accurate detection of individual flowers in crowded clusters is essential for autonomous pollinators.
• Yield Estimation: Counting flowers in dense clusters allows for more accurate crop yield predictions.
• Stress Evaluation: Monitoring flower density and health in real-time via UAVs or field robots.
• Scalability: The study demonstrates that by selecting the appropriate YOLO variant and annotation strategy, precision farming applications can be scaled effectively, balancing computational cost with detection accuracy.