MMdetection中对 multiscale技术的实现和使用

multiscale术语说明

首先看一下对 multiscale(multi-scale) 是指什么？

resizing images to different scales at each iteration.

核心代码实现：

在mmdet/datasets/pipelines/transforms.py文件中的Resize 类实现中，其核心函数为两个：random_select和random_sample，分别在设置data.train.multiscale_mode为value和range模式时调用。

class Resize(object):
	...
	def _random_scale(self, results): # 成员函数
		...
		elif self.multiscale_mode == 'range':  
		 scale, scale_idx = self.random_sample(self.img_scale)  
		elif self.multiscale_mode == 'value':  
		 scale, scale_idx = self.random_select(self.img_scale)
		 ...

	def __call__(self, results): # 类被调用时默认执行函数 
		...
		self._resize_img(results)  
		self._resize_bboxes(results)  
		self._resize_masks(results)  
		self._resize_seg(results)  
		return results
	...

使用与调用：

The term of "multiscale training" is adopted in many papers, which indicates resizing images to different scales at each iteration.
You only need to modify the data.train.img_scale field in the config file.
For example, you can use a list of scales [(1333, 800), (1666, 1000)] so that each image will randomly select a scale between (1333, 800) and (1666, 1000). In our paper, we use the setting [(400, 1600), (1400, 1600)] which means the short edge are randomly sampled from 400~1400, and the long edge is fixed as 1600.
from About multiscale training · Issue #78 · open-mmlab/mmdetection

![[Pasted image 20220323173932.png]]
![[Pasted image 20220323173953.png]]

在 swin-transformer 中的configs/swin/mask_rcnn_swin_tiny_patch4_window7_mstrain_480-800_adamw_1x_coco.py文件中使用：

![[Pasted image 20220323185912.png]]