one-stage anchor-free
These concepts come from one-stage anchor-free (not understood)
Questions:
one-stage / two-stage difference?
anchor-based / anchor-free difference?
Reference: 检测专题（一）：目标检测关键认知专题开篇了，跟Dr.He学检测，点赞关注不迷路_哔哩哔哩_bilibili
backbone
(Not understood, just copied down first)
The core function is to provide combinations of various receptive field sizes and center strides for detection, to satisfy the detection of targets of different scales and categories.
For example: resnet
- Each feature map inherently has semantic expression capability (what can be learned from the feature map for subsequent prediction, how well it can learn is predetermined — the receptive field on the feature map is already fixed, it cannot predict targets beyond the region)
- Each feature map provides an inherent downsampling factor, which to some extent determines the scale of targets the feature map can adapt to
Reference 检测专题二：backbone作为检测网络的核心组件之一，究竟有何关键作用？_哔哩哔哩_bilibili
neck
Accepts several feature maps from the backbone, then processes them (fuses features, enhances expression ability), outputs processed feature maps with the same width for the head to use.
For example: NaiveNeck, FPN, BiFPN, PANet, NAS-FPN
Reference 检测专题三：检测网络中neck的核心作用_哔哩哔哩_bilibili
head
???
For example: FCOS-Head
Reference 检测专题四：如何设计检测网络的head_哔哩哔哩_bilibili