What is object detection?

Object detection is a computer-vision task that predicts which object classes are present and locates individual instances, usually with bounding boxes and confidence scores.

How is object detection different from image classification?

Image classification assigns labels to an entire image. Object detection identifies and locates one or more individual objects, allowing a downstream system to reason about count and position.

How much training data does an object detector need?

There is no universal number. The requirement depends on class variability, pretrained-model fit, image conditions and the cost of errors. A representative pilot and learning-curve experiments provide better evidence than a generic target.

Which metrics should an object-detection project use?

Use per-class precision, recall and average precision alongside slice metrics, localization quality, end-to-end latency, resource use, human-review load, business outcomes and severe-error counts.

What is IoU in object detection?

Intersection over union measures overlap between predicted and ground-truth boxes by dividing their intersection area by their union area. Evaluation protocols use IoU thresholds to define adequate localization.

How should confidence thresholds be selected?

Select thresholds on representative validation data by balancing the real costs of false positives, false negatives and human review. High-impact classes may require separate thresholds and escalation rules.

Why can a detector perform well in testing but fail in production?

Common causes include data leakage, unrepresentative test images, changed cameras or lighting, preprocessing differences, pipeline latency, concept drift and evaluation that ignores the actual workflow.

Should object detection run at the edge or in the cloud?

Edge inference can reduce latency and data transfer, while cloud inference can simplify centralized scaling and updates. The choice depends on connectivity, privacy, hardware, workload, cost and recovery requirements.

How often should an object detector be retrained?

Retrain when monitored evidence shows meaningful drift, new classes or environments, corrected labelling rules or unacceptable performance. A calendar alone is not a sufficient retraining trigger.

What should be monitored after deployment?

Monitor input quality, camera availability, class and score distributions, end-to-end latency, alert volume, reviewed precision and recall, override patterns, business outcomes and policy incidents.

How should a team choose an object-detection algorithm?

Benchmark a small shortlist on representative data and target hardware against accuracy, latency, memory, maintainability and deployment constraints. See Rusaka's dedicated object-detection algorithm comparison for the model-family trade-offs.

Traditional Chinese · Rusaka

比較流行的物體偵測演算法：綜合指南

在電腦視覺領域，物件偵測是一項關鍵技術，使機器能夠識別和定位影像或影片中的物件。本文深入研究了流行的目標檢測演算法的比較，深入了解它們的工作原理和應用。