Skip to main content

Image Classification — Model Analysis

The Image Model Analysis tab loads a trained .pt or .onnx file and provides inference, validation, and chart tools. The page is split: left panel (controls + result tabs) and right panel (inference preview, 40% width).

Input Files

InputWhat to select
ModelBrowse to .pt (PyTorch) or .onnx trained model file
ImagesFolder of images to run inference on (does not need to be training data)
data.yamlProject YAML file (required for validation; optional for inference only)

Inference Configuration

ParameterDescriptionDefault
ConfConfidence threshold — detections below this score are discarded0.25
IoUNMS threshold — controls bounding box overlap before discarding0.45
ImgSzImage size for model input — must match training size640

Running Inference

  1. Select Model file and Images folder
  2. Adjust Conf, IoU, ImgSz if needed
  3. Click "▶ Run Inference"
  4. When complete, Prediction Table opens automatically

Prediction Table columns: Image (filename) · Detections (total objects) · Avg Conf (average confidence score).

NOTE

If Detections shows 0, check the confidence threshold is not too high, that the model was trained on the same classes, and that file selections are correct.

Inference Preview Panel

ActionHow
View a specific imageClick any row in the Prediction Table
Navigate imagesUse ◀ Prev / Next ▶ buttons or click table rows
ZoomScroll mouse wheel over the preview image
PanClick and drag the preview image
Full-screenDouble-click image; press Escape to close
Reset zoom"↺ Reset View" button in the Charts toolbar

Running Validation

  1. Select Model file and data.yaml
  2. Click "📊 Run Validation"
  3. When complete, Validation Metrics tab opens automatically
MetricWhat it means
mAP@0.50Mean Average Precision at IoU 0.50 — primary YOLO accuracy metric
mAP@0.50:0.95mAP averaged across IoU 0.50–0.95 — stricter measure
PrecisionFraction of detections that were correct (few false alarms)
RecallFraction of actual objects that were detected (few missed detections)
Per-class AP50AP@0.50 for each individual class — identifies weak classes

Charts Tab — PR Curve & F1 Curve

  • Precision-Recall (PR) Curve — trade-off between Precision and Recall at different confidence thresholds. Area under the curve = mAP score. A curve close to the top-right corner indicates a well-performing model.
  • F1-Confidence Curve — F1 score plotted against confidence threshold. The peak indicates the optimal Conf value for production use.
TIP

If the PR curve drops steeply at low recall, the model is missing many objects. If it drops at low precision, the model is generating false detections. Use this to decide on threshold tuning, more data collection, or a larger model variant.