物体检测

It differs from image classification, which can assign a label without locating the object, and segmentation, which describes pixel-level regions.

主要要点

• Define consistent instance annotations.
• Report matching and threshold settings.
• Test small, hidden, and crowded objects.

A detection dataset needs consistent labels and location annotations. Define how to handle partly hidden objects, very small instances, and ambiguous categories. Inconsistent boxes or omitted objects can confuse both training and evaluation. The model’s output usually includes a score and a location for each candidate. Postprocessing may remove overlapping duplicate predictions or apply a threshold. Those settings affect the balance between missed objects and false detections and should be recorded with the result. Evaluate localization and category correctness together. Intersection over union measures the overlap between a predicted region and a reference region. Precision and recall also depend on matching rules, score thresholds, and which object sizes are included. Test real capture conditions, including blur, lighting changes, occlusion, and crowded scenes. A detector can appear strong on large isolated objects while missing the small or partly hidden objects that matter in deployment. Define how uncertain detections are reviewed before they trigger consequential actions.

1. Use two invented 10-by-10 boxes. The second is shifted 5 units horizontally, so they overlap over a 5-by-10 region.
2. The intersection area is 50 and the union is 100+100−50 = 150. Intersection over union is 50/150, about 0.33.
3. Under a 0.5 matching threshold, the boxes would not count as a sufficient localization match despite substantial visible overlap.

The constructed geometry explains one evaluation component; it is not a detector benchmark.