karthi1008/Phygital_Interaction

Phygital Interaction

An Android application that uses AI-powered object detection to identify Kinder Joy toys in real time through a smartphone camera.

The project focuses on building a reliable on-device detection system using a custom-trained YOLO model and ONNX Runtime.

Simply point the camera at a supported Kinder Joy toy, and the application detects and identifies it.

---

📱 How It Works

Physical Kinder Joy Toy
          ↓
      Camera Input
          ↓
    Image Processing
          ↓
     YOLO AI Model
          ↓
    Object Detection
          ↓
 Confidence Validation
          ↓
     Toy Identified

The application continuously processes camera frames and checks whether a trained Kinder Joy toy is present.

The user does not need to manually select the toy. The application uses the camera and trained AI model to recognize it.

---

✨ Features

• 📷 Real-time camera detection
• 🤖 Custom-trained YOLO object detection model
• ⚡ ONNX Runtime inference on Android
• 🎯 Confidence-based detection filtering
• 🔲 Real-time bounding box detection
• 🧸 Multiple Kinder Joy toy classes
• 📱 Camera orientation handling
• ⚙️ On-device AI inference
• 📦 Optimized ONNX model support

---

🏗️ System Architecture

The application processes camera frames through the following pipeline:

Android Camera
      ↓
CameraX
      ↓
Image Preprocessing
      ↓
YOLO ONNX Model
      ↓
ONNX Runtime
      ↓
Detection Output
      ↓
Confidence Filtering
      ↓
Non-Maximum Suppression
      ↓
Bounding Box Mapping
      ↓
Toy Confirmed
      ↓
Tells The Name of The Toy

The model identify what the toy is in the camera frame.

---

🤖 AI Model

The project uses a custom-trained YOLO object detection model.

For each detected object, the model returns:

• Class name
• Confidence score
• Bounding box coordinates

Example:

Detected: Kinder Joy Toy

Confidence: 92%

Bounding Box:
X1: 120
Y1: 240
X2: 420
Y2: 640

The application filters low-confidence detections to reduce incorrect results.

---

📦 Model Export Comparison

Different ONNX export versions were tested to understand the trade-off between model size, performance, compatibility, and detection quality.

---

📊 Dataset Analysis

Before training, the dataset was analyzed to better understand the images, annotations, classes, and object distribution.

---

Class Distribution

This shows how the training data is distributed between the different toy classes.

---

Classes

The dataset contains multiple toy classes used for object detection.

---

Top Classes

This visualization shows the most represented classes in the dataset.

---

Split Distribution

This shows how the dataset is distributed across the different dataset splits.

---

Image Dimensions

The dataset contains images with different dimensions and aspect ratios.

---

Image File Size

This visualization shows the distribution of image file sizes.

---

Image Formats

This shows the image formats used in the dataset.

---

Annotation Locations

This visualization shows where objects are located throughout the training images.

---

Bounding Box Dimensions

The following analysis shows the distribution of bounding box sizes.

---

Bounding Box Dimensions — Additional Analysis

An additional bounding-box analysis is provided below.

---

Objects Per Image

This graph shows how many annotated objects are present in each image.

---

🧠 Training Analysis

The following images show the model training behaviour and optimization process.

---

Box Loss

Box loss measures the model's bounding-box localization error during training.

---

Classification Loss

Classification loss measures how well the model learns to distinguish between different toy classes.

---

DFL Loss

Distribution Focal Loss helps improve bounding-box localization.

---

Learning Rate

The learning-rate progression during training is shown below.

---

📈 Model Performance

The trained model was evaluated using standard object-detection metrics.

---

mAP@50

This measures detection accuracy at an IoU threshold of 0.50.

---

mAP@50–95

This is a stricter evaluation across multiple IoU thresholds.

---

Precision

Precision shows how many detected objects are actually correct.

---

Recall

Recall shows how successfully the model detects objects that are actually present.

---

🎯 Confidence Analysis

Every prediction produced by the AI model has a confidence score.

For example:

0.95 → Very confident
0.80 → Strong detection
0.60 → Possible detection
0.30 → Low confidence

Choosing the correct confidence threshold is important for avoiding false detections while still detecting the toy reliably.

---

F1-Confidence Curve

The F1 score helps find a balance between precision and recall at different confidence thresholds.

---

Precision-Confidence Curve

This shows how precision changes with different confidence thresholds.

---

Recall-Confidence Curve

This shows how recall changes with different confidence thresholds.

---

Precision–Recall Curve

This shows the relationship between precision and recall across different confidence levels.

---

🔍 Confusion Matrix

The confusion matrix shows how well the model distinguishes between the different toy classes.

---

Normalized Confusion Matrix

The normalized version makes it easier to compare the detection accuracy of each class.

---

💻 System Performance

The following images show resource usage and system behaviour during the experiment.

---

CPU & RAM Usage

---

GPU Utilization & Memory

---

GPU Temperature

---

Disk I/O

---

Network I/O

---

Run Information

---

System Information

```

The focus is on making object detection accurate, fast, and practical for running directly on a mobile device.

This project combines computer vision, mobile development, artificial intelligence to explore that idea.

---

Kinder Joy Toy Detection Dataset

Dataset Description

The Kinder Joy Toy Detection Dataset is a custom computer vision dataset created for detecting and identifying Kinder Joy toys from images using object detection models.

The dataset was developed as part of the Phygital Interaction project, which explores the connection between physical objects and digital experiences. The primary goal is to enable a mobile application to recognize Kinder Joy toys through a camera in real time.

Images in the dataset contain different Kinder Joy toys captured from multiple viewpoints and under varying conditions. Each object is annotated with a bounding box and its corresponding class label. The dataset can be used to train and evaluate object detection models such as YOLO and other compatible computer vision architectures.

• Curated by: Karthi
• Funded by [optional]: Independently developed / No external funding
• Shared by [optional]: Karthi
• Language(s) (NLP): Not applicable
• License: CC BY 4.0

Dataset Sources

• Repository: https://github.com/Karthi-1008/Phygital_Interaction
• Paper [optional]: Not available
• Demo [optional]: Not currently available

Uses

Direct Use

This dataset is intended primarily for:

• Training object detection models to recognize Kinder Joy toys.
• Developing real-time toy recognition applications.
• Testing mobile and edge-based computer vision models.
• Training YOLO-based object detection systems.
• Evaluating the effect of image resolution, model architecture, and quantization on object detection performance.
• Supporting physical-to-digital interaction experiments where a detected toy can trigger a corresponding digital experience.

Out-of-Scope Use

This dataset is not designed for:

• Facial recognition or person detection.
• General-purpose object detection.
• Identifying objects unrelated to the included Kinder Joy toy classes.
• Safety-critical or high-risk applications.
• Making decisions about individuals or groups.

The dataset may not perform reliably when used to detect toys that were not included in the training data or when images differ significantly from the conditions represented in the dataset.

Dataset Structure

The dataset consists of images and object detection annotations.

Each image may contain one or more annotated objects. Every annotation includes:

• Class label: The category or identity of the Kinder Joy toy.
• Bounding box: The location of the toy within the image.

The dataset is structured for use with object detection frameworks and may be exported in formats such as:

• YOLO
• COCO
• JSON
• CSV

The exact dataset split structure may include training, validation, and test subsets depending on the exported version.

Dataset Creation

Curation Rationale

The dataset was created to support the development of a real-time Kinder Joy toy detection system.

Existing general-purpose object detection datasets do not contain the specific Kinder Joy toy classes required for this project. Therefore, a custom dataset was necessary to train a model capable of distinguishing between the selected toys.

The dataset is designed to support efficient inference on mobile devices, where model size, speed, and detection accuracy are important constraints.

Source Data

The source data consists of images of Kinder Joy toys collected specifically for this project.

Images were selected to represent the target objects from different viewpoints and visual conditions to improve the robustness of the detection model.

Data Collection and Processing

Images were collected and prepared for object detection training.

The data preparation process included:

1. Capturing or collecting images of the target Kinder Joy toys.
2. Organizing images into relevant toy categories.
3. Annotating the location of each toy using bounding boxes.
4. Assigning a class label to each annotated object.
5. Exporting the annotated dataset into formats compatible with machine learning frameworks.
6. Preparing the dataset for training, validation, and testing.

Image annotation and dataset export were performed using an object detection dataset management platform.

The dataset may include variations in:

• Camera angle
• Object orientation
• Distance from the camera
• Background
• Lighting conditions
• Scale and position of the object within the image

Who are the source data producers?

The source images were collected specifically for this project by the dataset creator.

The images focus on physical Kinder Joy toys and are not intended to contain identifiable information about individuals.

Annotations

The dataset contains object detection annotations.

Each target toy is labeled with its corresponding class name and enclosed using a bounding box.

Annotation Process

Annotations were created manually using an object detection annotation platform.

The annotation workflow involved:

1. Opening each image in the annotation tool.
2. Drawing a bounding box around the visible toy.
3. Assigning the correct class label.
4. Reviewing annotations for incorrect labels or inaccurate bounding boxes.
5. Exporting the dataset in formats suitable for training object detection models.

The annotations were created specifically to support supervised object detection.

Who are the annotators?

The annotations were created and reviewed by the dataset creator as part of the Phygital Interaction project.

Personal and Sensitive Information

The dataset is not intended to contain personal, sensitive, or private information.

It focuses on images of physical Kinder Joy toys. No intentional collection of:

• Personally identifiable information
• Facial data
• Addresses
• Financial information
• Health information
• Political information
• Other sensitive personal data

was performed.

If any unintentionally captured identifying information is discovered, the affected image should be removed or appropriately anonymized before further distribution.

Bias, Risks, and Limitations

The dataset has several limitations.

Because it was created for a specific set of Kinder Joy toys, the trained model may have reduced performance when detecting:

• Toys not included in the dataset.
• New versions or visually different variants of existing toys.
• Toys that are heavily occluded.
• Objects under extreme lighting conditions.
• Very small or distant objects.
• Images with significant motion blur.
• Backgrounds or environments not sufficiently represented in the training data.

Dataset imbalance between classes may also affect model performance. If some toy categories contain significantly more images than others, the trained model may perform better on those categories.

The dataset should therefore be considered a project-specific dataset rather than a general benchmark for universal toy detection.

Recommendations

Users should evaluate model performance on their own target environment before deploying a model trained on this dataset.

For improved robustness, users are encouraged to:

• Collect additional images representing real deployment conditions.
• Maintain a balanced number of samples across classes.
• Include multiple camera angles and object orientations.
• Include different backgrounds and lighting conditions.
• Validate annotations before training.
• Evaluate the model using separate validation and test data.
• Retrain or fine-tune the model when introducing new toy classes.

Citation

If you use this dataset, please cite the associated project repository.

BibTeX:

@misc{kinder_joy_toy_detection_dataset,
  author = {Karthi},
  title = {Kinder Joy Toy Detection Dataset},
  year = {2026},
  howpublished = {\url{https://github.com/Karthi-1008/Phygital_Interaction}},
  note = {Dataset created for the Phygital Interaction project}
}

APA:

Karthi. (2026). Kinder Joy Toy Detection Dataset. Phygital Interaction Project. https://github.com/Karthi-1008/Phygital_Interaction

Glossary

• Object Detection: A computer vision task that identifies objects and determines their location within an image.
• Bounding Box: A rectangular region used to indicate the location of an object.
• Class Label: The category assigned to an annotated object.
• YOLO: A family of real-time object detection models.
• Inference: The process of using a trained model to make predictions on new input data.
• Phygital Interaction: Interaction that connects physical objects with digital experiences.

More Information

This dataset is part of the Phygital Interaction project.

The project focuses on recognizing physical Kinder Joy toys through computer vision and using the detection result to enable corresponding digital interactions.

Project repository:

https://github.com/Karthi-1008/Phygital_Interaction

Dataset Card Authors

Karthi

Dataset Card Contact

For project information, updates, or questions, please refer to the GitHub repository:

https://github.com/Karthi-1008/Phygital_Interaction

Also available at Kaggle

https://www.kaggle.com/datasets/karthikeyan100/kinder-joy-toys