Table 2. Detailed information about the proposed network about lffd-a-light-and-fast-face-detector-for-edge-devices HOT 4 CLOSED

@sc199505
RF / ( (upper_bound+lower_bound)/2 )
for example: RF:55, scale [10, 15] -> ratio = 55 / ( (10+15)/2 ) = 4.4

from lffd-a-light-and-fast-face-detector-for-edge-devices.

您好，谢谢您的回答，对于Table2 还有点不明白，论文里“For an instance, RFs of 100 pixels are able to predict faces between 20 pixels to 40 pixels.”是不是可以这样理解为对于实际感受野其范围是感受野的0.2倍到0.4倍之间，那么对于Table2里的Continuous face scale是不是可以由计算感受野的0.20.4计算得来，但是这样算下来:
RF = 55 0.2 * RF = 11 0.4 * RF = 22 ; 1122
RF = 71 0.2 * RF = 14.2 0.4 * RF = 28.4 ; 14.228.4
RF = 111 0.2 * RF = 22.2 0.4 * RF = 44.4; 22.244.4
RF = 143 0.2 * RF = 28.6 0.4 * RF = 57.2; 28.657.2
RF = 223 0.2 * RF = 44.6 0.4 * RF = 89.2; 44.689.2
RF = 383 0.2 * RF = 76.6 0.4 * RF = 153.2; 76.6153.2
RF = 511 0.2 * RF = 102.2 0.4 * RF = 204.2; 102.2204.2
RF = 639 0.2 * RF = 127.8 0.4 * RF = 255.6; 127.8~255.6
和Table2里的Continuous face scale明显不匹配，是不是我的理解有问题啊，望指教。

from lffd-a-light-and-fast-face-detector-for-edge-devices.

@sc199505
具体来说,这里并没有一个定量的具体对应关系.
文中有相关的解释,这个设置(不同感受野针对不同尺度范围)会和具体的任务挂钩.
文中的backbone版本就如Table 2所示. repo中我们提供了v2的版本, 这个版本稍微简化了设计, 详情请参考/face_detection/symbol_farm/symbol_structures.xlsx.
文章的核心就是在讨论感受野应该如何去配合待检尺寸范围, 从而设计出更简洁的backbone.

from lffd-a-light-and-fast-face-detector-for-edge-devices.

好的，谢谢您的回答。

from lffd-a-light-and-fast-face-detector-for-edge-devices.