The contents in this post are excerpted from the paper “YOLOv3: An Incremental Improvement” ¹, with a little bit modification, as my notes for this paper.

Also refer to:

• yolo系列之yolo v3【深度解析】
• Yolov3相关算法的原理及实现

2. The Deal

2.1. Bounding Box Prediction

Use anchors.

2.2. Class Prediction

We use independent logistic classifiers to predict classes instead of a softmax. In another word, we use multilabel classification instead of multiclass classification. During training we use binary cross-entropy loss.

In this dataset there are many overlapping labels (i.e. Woman and Person). Using a softmax imposes the assumption that each box has exactly one class which is often not the case. A multilabel approach better models the data.

2.3. Predictions Across Scales

YOLOv3 predicts boxes at 3 different scales. Our system extracts features from those scales using a similar concept to feature pyramid networks (FPN).

2.4. Feature Extractor

Use Darknet-53.

Darknet-53 has similar performance to ResNet-152 and is 2 times faster. ResNets have too many layers and are not very efficient.

3. How We Do

With FPN, YOLOv3 has relatively high \(\text{AP}_S\) performance, i.e. it is good at detecting small objects. However, it has comparatively worse performance on medium and larger size objects.

Reference: