Manual evaluation deliberately inspects model behavior on selected inputs alongside automated validation. It exposes unexpected outputs during development and debugging, verifies specific test cases, demonstrates capabilities to stakeholders, and builds confidence before operational use. section 8.1.7
Use manual inspection to:
- Verify behavior on deliberately selected or previously unseen test images. section 8.1.7
- Compare predicted objects with the intended targets and investigate unexpected classifications. section 8.1.7
- Interpret confidence scores, where values near 1 indicate stronger detection confidence and values near 0 indicate weaker confidence. section 8.1.7
Inspecting object-detection outputs
Run inference across images in a samples directory, display each annotated result, and pause for inspection. A generic YOLO model may detect faces as persons and misclassify an ID card as a low-confidence book; repeating the inspection with trained best weights reveals whether training improves the result. section 8.1.7
Listing 8.15 Running inference with the default Nano model
import glob
import cv2
from ultralytics import YOLO
model = YOLO("yolov8n.pt") #1
for file in glob.glob("samples/**.jpg"): #2
result = model(cv2.imread(file))
res_plotted = result[0].plot() #3
cv2.imshow("result", res_plotted) #4
cv2.waitKey(0)
Provides a runnable inspection loop using Ultralytics YOLO, OpenCV, sample images, plotted detections, and interactive pauses. section 8.1.7
