Image Classification

IMPORTANT

Every change in this tab must be saved by clicking 💾 Save Configuration before starting training. Training uses the last saved settings.

Model Variant

Variant	Speed / Accuracy	Best Use Case
YOLOv8n (Nano)	Fastest / Lower accuracy	Edge devices with very limited compute
YOLOv8s (Small)	Fast / Moderate	General-purpose edge deployment
YOLOv8m (Medium)	Moderate / High	Powerful edge devices or desktop inference
YOLOv8l (Large)	Slower / Highest	Maximum accuracy where compute is available

TIP

Start with YOLOv8n. Upgrade to a larger variant only if accuracy is insufficient.

Task Type

Task Type	Description
Detection	Outputs bounding boxes with class labels and confidence scores

Dataset Format

Format	Description
YOLO	YOLO-format .txt label files — default and recommended
COCO	COCO JSON annotation format
Pascal VOC	XML annotation format

Image Configuration

Parameter	Description	Recommended
Image Size	Resolution images are resized to before training. Larger = better for small objects, slower.	640 px
Channels	RGB (3-channel) or Grayscale (1-channel)	RGB
Train Split	Percentage of dataset used for training	80%
Validation Split	Percentage reserved for validation metrics	20%

Data Augmentation

Augmentation	Effect	Recommended
Horizontal Flip	Mirrors image left-to-right randomly	Yes
Rotation	Random angle rotation	Yes
Brightness	Random brighten or darken	Yes
Mosaic	Combines 4 training images into one tile — very effective for detection	Yes
Blur	Applies Gaussian blur	Optional
Noise	Adds random pixel noise	Optional

Inference Configuration

Parameter	Description	Default
Confidence Threshold	Minimum confidence score for a detection to be returned	0.25
IoU Threshold	NMS threshold for bounding box overlap before discarding lower-confidence boxes	0.45
Max Detections	Maximum bounding boxes returned per image	100

Training Hyperparameters

Parameter	Description	Default
Epochs	Complete passes through the training dataset	100
Patience	Epochs without improvement before early stopping	4
Batch Size	Images per training step. Reduce if out of memory.	16

NOTE

QAT (Quantization Aware Training) is hidden for Image projects — YOLO uses a different quantization flow during ONNX export.

Dataset Splitting & YAML Generation

Before training, split your dataset and generate data.yaml via the Preprocessing tab.

1. Click "Split & Analyze Dataset" to scan datasets/, detect imbalances, and view statistics
2. Set Train Split (80%) and Validation Split (20%), then click the split button
3. Click "Generate YAML Files" — creates data.yaml in your project root

The split keeps images in their original class folders and only distributes label files into train/val subfolders:

datasets/
  human_face/
    image1.jpg
    image2.jpg
    labels/
      image1.txt
      image2.txt
    classes.txt
  labels/
    train/   ← training label .txt files
    val/     ← validation label .txt files

Typical data.yaml content:

path: /path/to/your/project
train: datasets/images/train
val: datasets/images/val
names:
  0: class_name_1
  1: class_name_2

TIP

After generating the YAML file, verify class names are correctly mapped and match the class IDs in your label files (0-indexed).