输入图像1920x1080,batch_size=8为例.

训练阶段

loss_dict = model(images,targets)

• 入参
  • images: List(Tensor(3,1920,1080))[8]
  • targets: List(dict()[3])[8] dict详情见下表:

key	type	dtype	size	remark
boxes	Tensor	float32	(n,4)1	the ground-truth boxes in [x1, y1, x2, y2] format, with 0 <= x1 < x2 <= W and 0 <= y1 < y2 <= H.
labels	Tenosr	int64	(n,)	the class label for each ground-truth box
maskes	Tensor	uint8	(n,1920,1080)[N,H,W]	the segmentation binary masks for each instance,实际就是0和1,有对象的区域就是1,否则就是0,这个照片有多少个对象就有多少个mask
area*	Tensor	float32	(n,)	对象面积
iscrowd*	Tensor	int64	(n,)	是否为一群对象(coco数据集会标注)
image_id*	int			图像编号

*为非必要参数,有一些数据集合处理的时候会标注上去*

• 返回 loss_dict dict()[5] dict详情见下表:

key	type	dtype	size	损失函数	remark
loss_classifer	Tensor	float32	()	CrossEntropyLoss	对象分类损失（Classification Loss）：
loss_box_reg	Tensor	float32	()	Smooth L1 Loss/MSE	边界框回归损失（Bounding Box Regression Loss）：
loss_mask	Tensor	float32	()	Binary Cross-Entropy Loss	掩膜损失（Mask Loss）：
loss_objectness	Tensor	float32	()	CrossEntropyLoss	RPN分类损失（RPN Classification Loss）：前景/背景二分类损失
loss_rpn_box_reg	Tensor	float32	()	Smooth L1/MSE	RPN边界框回归损失（RPN Bounding Box Regression Loss）

推理阶段

predict = model(images)

• 入参 images: List(Tensor(3,1920,1080))[8] 不变
• 返回 predict:List(dict()[4])[8] dict详情见下表:

key	type	dtype	size	remark
boxes	Tensor	float32	(m,4)2	the predicted boxes in [x1, y1, x2, y2] format,预测的所有的边界框
labels	Tensor	int64	(m,)	the predicted labels for each instance
boxes	Tensor	float32	(m,)	the scores or each instance
boxes	Tensor	float32	(m,1,1920,1080)[M, 1, H, W]	the predicted masks for each instance, in 0-1 range. In order to obtain the final segmentation masks, the soft masks can be thresholded, generally with a value of 0.5 (mask >= 0.5).实际存储的是一个软掩膜,0.5以下的也有,存在比较平滑的过度

参考官方文档
maskrcnn_resnet50_fpn

---

1. n为此图像上的实例个数 ↩︎

2. m为此图像 ↩︎