Hey there! Since you're new to computer vision and machine learning, let's break down these two questions about YOLOv3's anchor boxes clearly—no jargon overload, promise.

1. Why did modifying anchor boxes for face detection tank your results? And how important are anchors for category prediction? #

First, let's clarify: anchor boxes don't directly handle category prediction in YOLOv3, but they have a huge indirect impact on how well your model performs overall, including category accuracy. Here's why:

• Anchors act as shape priors for your model. YOLOv3 uses anchors to define "default" box shapes that it expects to see in your data. For face detection, most faces have a roughly square aspect ratio (like 1:1 or 1.2:1). If you swapped those out for anchors that are super wide (like 4:1) or tall (like 1:4), the model can't find a good starting point to predict the actual face bounding boxes.
• When your anchors don't match your target shapes, the model struggles to learn the correct offset values (the adjustments it makes to anchors to fit the real object). Even if it can correctly classify a face, if the bounding box is way off, the detection will be rejected (thanks to IoU thresholds) or look useless. That's why your results got so bad.
• To put it simply: anchors give the model a head start. Without the right priors, the model has to learn both the shape of your targets and their categories from scratch, which is way harder—especially if you're working with a smaller dataset.

2. How do YOLOv3's 9 default anchors handle 80 diverse object shapes? #

The 9 default anchors (3 per feature scale) aren't tied to specific categories—they're clustered based on the aspect ratios and sizes of all object bounding boxes in the COCO dataset (the dataset YOLOv3 was trained on). Here's the breakdown:

• The clustering uses k-means (with a custom distance metric that prioritizes IoU over Euclidean distance) to group all COCO's object boxes into 9 clusters. These clusters represent the most common shapes and sizes across all 80 classes.
• The anchors are split across 3 feature scales:
  • Small anchors (like (10,13), (16,30), (33,23)) are used on the highest-resolution feature map, which detects small objects (like birds, stop signs).
  • Medium anchors (like (30,61), (62,45), (59,119)) go on the middle-resolution map for medium-sized objects (like dogs, chairs).
  • Large anchors (like (116,90), (156,198), (373,326)) are used on the lowest-resolution map for big objects (like cars, people standing full-body).
• Even though the 80 classes have varied shapes, these 9 anchors cover the majority of their bounding box distributions. The model learns to adjust each anchor's position and size (via offset predictions) to fit whatever object it's detecting, regardless of category.

If you're working with a dataset that's very different from COCO (like your face detection project), you should always re-cluster anchors specific to your target objects—this will give your model the right priors to work with.

内容的提问来源于stack exchange，提问作者nOpe