学习资源
Paddle官方教程: 《AI快车道PaddleDetection – 课节4: 闪电版目标检测算法PP-PicoDet》
PicoDet增强版官方介绍: 超强目标检测算法矩阵（PicoDet-XS）
PicoDet论文: PP-PicoDet: A Better Real-Time Object Detector on Mobile Devices
关键点视频也需要学习一下；
1 模型介绍
论文阅读笔记：《PicoDet论文译读笔记》
模型类型: Anchor-Free
关于PicoDet的整体描述，我觉得可以参照VFNet论文对这类模型的介绍：
[VFNet Paper]: Another type of anchor-free detectors are similar to anchor-based one-stage methods, but they remove the usage of anchor boxes. Instead, they classify each point on the feature pyramids [25_FPN] into foreground classes or background, and directly predict the distances from the foreground point to the four sides of the ground-truth bounding box, to produce the detection. Popular methods include DenseBox [26_DenseBox], FASF [27_FASF], FoveaBox [15_FoveaBox], FCOS [9_FCOS], and SPAD [28_SPAD].
技术一览表：
| 名称 |
cfg |
| 参数初始化 |
KaimingNormal |
| Loss |
l
o
s
s
v
f
l
+
2
⋅
l
o
s
s
g
i
o
u
+
0.25
⋅
l
o
s
s
d
f
l
loss_{vfl}+2\cdot loss_{giou} + 0.25\cdot loss_{dfl}
lossvfl+2⋅lossgiou+0.25⋅lossdfl |
2 损失函数
首先，来看看loss函数的整体公式：
l
o
s
s
=
l
o
s
s
v
f
l
+
2
⋅
l
o
s
s
g
i
o
u
+
0.25
⋅
l
o
s
s
d
f
l
loss = loss_{vfl}+2\cdot loss_{giou} + 0.25\cdot loss_{dfl}
loss=lossvfl+2⋅lossgiou+0.25⋅lossdfl
其中，
l
o
s
s
v
f
l
loss_{vfl}
lossvfl：Varifocal Loss。
l
o
s
s
g
i
o
u
loss_{giou}
lossgiou：GIoU Loss。
l
o
s
s
d
f
l
loss_{dfl}
lossdfl：Distribution Focal Loss。
2.1 Varifocal Loss —— 分类损失
Varifocal Loss是在VarifocalNet[paper]中提出的；
（关于VarifocalNet论文，请参见《【CVPR 2021】VarifocalNet: An IoU-aware Dense Object Detector的译读笔记》）
请参考《mmdetection最小复刻版(十六)：iou感知VarifocalNet深入分析》；
2.2 GIoU Loss —— 回归损失
用来回归边界框的坐标值；
2.3 Distribution Focal Loss —— 分类损失
Distribution Focal Loss出自论文《Generalized Focal Loss: Learning Qualified and Distributed Bounding Boxes for Dense Object Detection》；
参考资料：
DFLoss代码实现：
我看了一下PaddleDetection的实现跟MMDetection的实现几乎是一样的；
3 训练自定义数据
训练命令：
python tools/train.py -c configs/model/model_coco.yml
