How to modify the size of anchor?

[xyw7]

Hi mvoelk,

Thanks for your excellent work.

About the TBPP_train code, I'm confused that how do you generate anchor boxes on different feature maps. I didn't find an explicit method in your code about designing thoes boxes. I've looked through other version about SSD, and all of they almost provide a custom way to modify the size of anchor box like a config file or something. And the reason why I concern this is in my project, I need to detect a whole line text. And the aspect ratio has a lot of variation from 5 to 20, almost evenly distributed. So for a better perfomance on both precision and recall, I may need to modify the anchor size frequently. And about this issue, could you give me some advice? Any help will be appreaticed !

Best,
Xiyuan

[mvoelk]

aspect_ratio and scale in ssd_utils.py PriorUtil and PriorMap...

You can also set the values as attributes of the model object

ssd_detectors/tbpp_model.py

Line 101 in 257ae27

model.aspect_ratios = [[1,2,3,5,1/2,1/3,1/5] * 2] * num_maps

The PriorMap class has an method plot_boxes for visualizing the prior boxes.

[mvoelk]

In general, it makes sense to think about the receptive field when designing the prior boxes.

[mvoelk]

Why don't you use SegLink, you don't have the issue with the aspect ratio?

[xyw7]

Why don't you use SegLink, you don't have the issue with the aspect ratio?

Thanks for replying!

I just want to evaluate all classic text detect methods like TB, TB++ and Seglink etc. I want to find an algorithm that is balanced in accuracy and speed. And yeah I'll try anchor free model next.