Train Your Object Detection Model

Train a model to detect and locate objects in new images.

Training Configuration

Start Training

Navigate to Training > New Job
Select your dataset
Configure training:
- Model architecture: YOLO v4/v8, Faster R-CNN
- Base model: Pre-trained weights
- Epochs: 50-300 depending on dataset size
- Batch size: 8-32 based on GPU memory
Click Start Training

Architecture Options

Architecture	Speed	Accuracy	Best For
YOLO v4	Fast	Good	Real-time applications
YOLO v8	Faster	Better	Production deployment
Faster R-CNN	Slower	Higher	Maximum accuracy

from seeme import Client

client = Client()

# Create training job
job = client.create_training_job(
    name="Vehicle Detection v1",
    dataset_id=dataset.id,
    version_id=version.id,

    # Model configuration
    model_type="object_detection",
    architecture="yolov4",

    # Training parameters
    config={
        "epochs": 100,
        "batch_size": 16,
        "learning_rate": 0.001,
        "image_size": 640,

        # Data augmentation
        "augmentation": {
            "flip_horizontal": True,
            "flip_vertical": False,
            "rotation": 15,
            "scale": [0.8, 1.2],
            "mosaic": True
        },

        # Advanced
        "pretrained": True,
        "freeze_backbone": False,
        "warmup_epochs": 3
    }
)

print(f"Training job started: {job.id}")

# Create training job for object detection
curl -X POST "https://api.seeme.ai/api/v1/jobs" \
  -H "Authorization: myusername:my-api-key" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Vehicle Detection v1",
    "dataset_id": "your-dataset-id",
    "version_id": "your-version-id",
    "model_type": "object_detection",
    "architecture": "yolov4",
    "config": {
      "epochs": 100,
      "batch_size": 16,
      "learning_rate": 0.001,
      "image_size": 640,
      "augmentation": {
        "flip_horizontal": true,
        "rotation": 15,
        "mosaic": true
      },
      "pretrained": true,
      "warmup_epochs": 3
    }
  }'

# Get job status
curl -X GET "https://api.seeme.ai/api/v1/jobs/{job_id}" \
  -H "Authorization: myusername:my-api-key"

Monitor Training Progress

import time

# Monitor training progress
while job.status in ["pending", "running"]:
    job = client.get_training_job(job.id)

    if job.metrics:
        print(f"Epoch: {job.metrics.get('epoch', 0)}")
        print(f"Box Loss: {job.metrics.get('box_loss', 'N/A'):.4f}")
        print(f"mAP@0.5: {job.metrics.get('map_50', 'N/A'):.4f}")
        print(f"mAP@0.5:0.95: {job.metrics.get('map', 'N/A'):.4f}")
        print("---")

    time.sleep(30)

print(f"Training complete: {job.status}")

# Get job status and metrics
curl -X GET "https://api.seeme.ai/api/v1/jobs/{job_id}" \
  -H "Authorization: myusername:my-api-key"

# Get metrics history
curl -X GET "https://api.seeme.ai/api/v1/jobs/{job_id}/metrics" \
  -H "Authorization: myusername:my-api-key"

# Get trained model after completion
curl -X GET "https://api.seeme.ai/api/v1/models/{model_id}" \
  -H "Authorization: myusername:my-api-key"

# Run inference on test image
curl -X POST "https://api.seeme.ai/api/v1/inferences/{model_id}" \
  -H "Authorization: myusername:my-api-key" \
  -F "file=@./test_image.jpg"

Object Detection Metrics

Metric	Description	Target
mAP@0.5	Average precision at 50% IoU	> 0.7
mAP@0.5:0.95	Average across IoU thresholds	> 0.5
Precision	How many detections are correct	> 0.8
Recall	How many objects are found	> 0.8
F1 Score	Balance of precision/recall	> 0.75

IoU (Intersection over Union)

Predicted box: ┌──────┐
               │ ████ │ ← Overlap (intersection)
Actual box:    │ ████ ┐
               └──────┘ │
                        │
               └────────┘

IoU = Intersection Area / Union Area

Training Results

View Results

# Get trained model
model = client.get_model(job.model_id)

# View per-class performance
for cls in model.class_metrics:
    print(f"{cls['name']}: AP={cls['ap']:.3f}")

# View confusion matrix
confusion = model.confusion_matrix

Test on Sample Images

# Run inference on test images
results = client.predict(
    model_id=model.id,
    item="./test_image.jpg"
)

for detection in results.inference_items:
    print(f"Found {detection.prediction} at ({detection.x}, {detection.y})")
    print(f"  Size: {detection.width}x{detection.height}")
    print(f"  Confidence: {detection.confidence:.2%}")

Hyperparameter Tuning

Key Parameters

Parameter	Effect	Typical Range
Learning rate	Training speed/stability	0.0001 - 0.01
Batch size	GPU utilization, generalization	8-64
Image size	Detail vs speed	320-1280
Confidence threshold	Precision vs recall	0.25-0.7
NMS threshold	Duplicate filtering	0.3-0.6

Auto-Tune

# Run hyperparameter search
search = client.create_hyperparam_search(
    dataset_id=dataset.id,
    search_space={
        "learning_rate": [0.001, 0.0001, 0.00001],
        "batch_size": [8, 16, 32],
        "image_size": [416, 640]
    },
    metric="map_50",
    max_trials=10
)

Best Practices

Start with pre-trained weights - Faster convergence
Use augmentation - Especially mosaic for detection
Monitor mAP, not just loss - Loss can decrease while mAP plateaus
Balance classes - Oversample rare classes if needed
Test early - Run inference after 50% training to catch issues

Troubleshooting

Problem	Likely Cause	Solution
Low mAP	Insufficient data	Add more images, use augmentation
High loss, no improvement	Learning rate too high	Reduce learning rate
Detects wrong classes	Label inconsistency	Review and fix annotations
Misses small objects	Image size too small	Increase image_size
Many false positives	Low quality annotations	Review annotation quality

Next Step

4. Deploy →

Annotate Objects Deploy Your Object Detection Model

Train Your Object Detection Model

Training Configuration

Start Training

Architecture Options

Monitor Training Progress

Training Dashboard

Object Detection Metrics

IoU (Intersection over Union)

Training Results

View Results

Test on Sample Images

Hyperparameter Tuning

Key Parameters

Auto-Tune

Best Practices

Troubleshooting

Next Step