Tutorial: Training a Custom Object Detection Model

This tutorial guides you step by step through the process of training an object detection model tailored to your specific needs. You will learn how to prepare your data, configure the model, launch training, and evaluate the results.

Prerequisites

Before you begin, make sure you have the following:

• A Techsolut account with access to training features
• A set of images representative of your use case
• A basic understanding of computer vision concepts
• At least 50 images for initial results (more is preferable)

Step 1: Data Preparation

1.1 Image Collection

Gather images that well represent your usage scenario:
- Capture objects from different angles, lighting, and contexts
- Include images where objects are partially visible or slightly blurred
- Ensure sufficient resolution (minimum 640×640 pixels recommended)
- Vary backgrounds and conditions for better generalization

1.2 Data Organization

1. Create a new project by clicking "New" in the main menu
2. Select "Object Detection Project" as the type
3. Give your project a descriptive name
4. Import your images using the "Import Images" button
5. Organize them into training (70%), validation (15%), and test (15%) sets
6. Use the "Automatic Distribution" function or do it manually

1.3 Image Annotation

1. Access the annotation tool by clicking on the "Annotation" tab
2. Create the object classes you want to detect
3. Click on "Manage Classes" then "Add Class"
4. Give a clear name and choose a distinctive color for each class
5. Annotate your images by drawing bounding boxes around each object
6. Select the appropriate class before drawing each box
7. Ensure the box encompasses the entire object but remains as tight as possible
8. Remember to annotate all relevant objects in each image

Tip: Use keyboard shortcuts to speed up annotation. Press "?" to display the list of available shortcuts.

1.4 Annotation Quality Verification

1. Use the "Annotation Verification" tool to spot potential errors
2. Check class statistics to ensure a reasonable balance
3. Correct inconsistent or imprecise annotations
4. Confirm all images are annotated before moving to the next step

Step 2: Training Configuration

2.1 Base Model Selection

1. Go to the "Training" tab of your project
2. Click "New Model" to begin configuration
3. Choose an architecture suited to your use case:
4. YOLOv8-s: Balance between speed and accuracy, good general starting point
5. YOLOv8-n: Lighter version for deployment on resource-limited devices
6. YOLOv8-m/l: Larger versions for maximum accuracy
7. Faster R-CNN: Alternative if accuracy takes precedence over speed
8. EfficientDet: Good compromise between efficiency/accuracy for objects of varying sizes

2.2 Hyperparameter Configuration

Adjust training parameters according to your needs:

• Input size: 640×640 recommended to balance details and performance
• Batch size: 16-32 for most cases (reduce if memory issues)
• Number of epochs:
• 50-100 for initial training
• 10-30 for fine-tuning from a pre-trained model
• Learning rate:
• Initial: 0.01 (adjust to 0.001 for fine-tuning)
• Use the "cosine" scheduler with warm-up
• Data augmentation:
• Enable horizontal flip, slight rotations, and brightness adjustments
• Use mosaic and mixup for smaller datasets

2.3 Advanced Configuration

For experienced users, additional options are available:

• Optimizer: SGD or Adam (Adam often faster but may overfit)
• Loss function: CIoU by default (try DIoU for small or dense objects)
• EMA (Exponential Moving Average): Enable for more stable convergence
• Callbacks: Configure metric logging and saves
• Multi-scale training: Enable to improve robustness to size variations

Step 3: Training Launch

3.1 Resource Preparation

1. Check available resources in the "Resources" tab
2. Reserve a GPU if available (strongly recommended)
3. Estimate training time based on your parameters and data

3.2 Starting Training

1. Check your configuration one last time in the summary page
2. Click "Start Training"
3. Follow initialization and confirm training starts without errors

3.3 Training Monitoring

During training, closely monitor these metrics:

• Training loss: Should decrease steadily
• Validation loss: Should follow the training loss trend
• mAP (mean Average Precision): Main performance metric, should increase
• Precision and recall: Balance between correct detections and false positives/negatives
• PR curves: Visualization of performance by class

Caution: If you observe that the training loss continues to decrease but the validation loss increases, you might be facing overfitting. Consider stopping training or adjusting parameters.

Step 4: Evaluation and Improvement

4.1 Results Analysis

Once training is complete:

1. Examine final metrics in the "Results" tab
2. Analyze performance by class to identify strengths and weaknesses
3. Visualize predictions on the test set for qualitative evaluation
4. Identify problematic cases and typical errors

4.2 Problem Diagnosis

If results are not satisfactory, identify the likely cause:

• Low precision but high recall: The model generates too many false alarms

• Solution: Increase confidence threshold or improve diversity of negative examples

• Low recall but high precision: The model misses correct detections

• Solution: Reduce confidence threshold or add more varied positive examples

• Low precision and recall: Fundamental training problems

• Solution: Check annotation quality, increase data quantity, adjust architecture

• Uneven performance between classes: Data imbalance or variable complexity

• Solution: Balance classes, add more examples for problematic classes

4.3 Iteration and Improvement

1. Adjust your approach based on diagnosis:
2. Collect more data if necessary
3. Refine existing annotations
4. Adjust hyperparameters

5. Try another architecture

6. Launch a new training with modifications

7. Compare results with previous training
8. Repeat until satisfactory performance is achieved

4.4 Model Export

Once satisfied with performance:

1. Select the best checkpoint in the training history
2. Click "Export Model"
3. Choose the appropriate format according to your use case:
4. ONNX: For maximum compatibility
5. TorchScript: For deployment with PyTorch
6. TensorRT: For high-performance inference on NVIDIA GPU
7. TFLite: For mobile and embedded devices

Conclusion

Congratulations! You now have a custom object detection model ready for deployment. You can use it directly in the Techsolut platform or export it to integrate into your own applications.

For advice on deployment and optimization, check out our tutorial "Model Optimization for Deployment".

Feel free to share your results and questions on our community forum to get feedback and suggestions from other users.