search

sfzhang15 / RefineDet

Single-Shot Refinement Neural Network for Object Detection, CVPR, 2018
Other
1.43k stars 393 forks source link

Questions about multi-scale test #180

Closed WangTianYuan closed 5 years ago

WangTianYuan commented 5 years ago

Hello, I've tried your multi-scale test code, but I have some questions: I test the single-scale test on M2Det, it can achieve the original mAP:34.6. But when I use 1 and 1.2x scale, I only got 31.0 mAP. In this case, I've chage the pribox config for 1.2x scale, so why I got less mAP? the result is: image image

sfzhang15 commented 5 years ago

@WangTianYuan What does "chage the pribox config for 1.2x scale" mean?

WangTianYuan commented 5 years ago

@sfzhang15 Sorry, it's prior_box, for input size 320, the default config is image so for 1.2x scale test, I have to change it like this: image Is this correct? I find that min_sizes and steps effect the result

sfzhang15 commented 5 years ago

@WangTianYuan The priorbox must not be modified during the multi-scale testing.

WangTianYuan commented 5 years ago

@sfzhang15 But for orginal scale 320, the feature map is [40, 20, 10, 5, 3, 1]. for 1.2 x scale, the feature map should be [48, 24, 12, 6, 3, 1]if we still use the default config, this exception will appear: The size of tensor a (100354) must match the size of tensor b (32756) at non-singleton dimension 0

sfzhang15 commented 5 years ago

@WangTianYuan The steps, min_sizes, max_sizes must not be modified during the multi-scale testing.