search

yqyao / FCOS_PLUS

Some improvements (center sample) about FCOS (FCOS: Fully Convolutional One-Stage Object Detection).
Other
321 stars 46 forks source link

How to set parameter "radius" of center sample #13

Open 18804601171 opened 4 years ago

yqyao commented 4 years ago

You can add it in config file (configs/fcos/fcos_R_50_FPN_1x_center_giou.yaml) @18804601171

18804601171 commented 4 years ago

@yqyao gt = gt[None].expand(K, num_gts, 4),in this line have a error,RuntimeError: expand(torch.cuda.FloatTensor{[1, 8525, 70, 4]}, size=[8525, 8525, 4]): the number of sizes provided (3) must be greater or equal to the number of dimensions in the tensor (4) ,expand shouldn't have error,why do you know

18804601171 commented 4 years ago

in addition,how to set the value of radius is suitable?

yqyao commented 4 years ago

In my experiment, 1.5 is the best. @18804601171

18804601171 commented 4 years ago

@yqyao why expand gt's dim,why not use original dim

yqyao commented 4 years ago

Gt will be implicitly expanded in this line, I just expanded it in advance. @18804601171 https://github.com/yqyao/FCOS_PLUS/blob/0d20ba34ccc316650d8c30febb2eb40cb6eaae37/maskrcnn_benchmark/modeling/rpn/fcos/loss.py#L66

18804601171 commented 4 years ago

How many dim of gt are passed into function get sample region @yqyao

yqyao commented 4 years ago

I don't understand your question very well, I think all gt will be passed into function to get sample region. Maybe you need to read our recent paper https://arxiv.org/abs/1912.02424. @18804601171

18804601171 commented 4 years ago

@yqyao The size of radius only determines the sampling range. If it is beyond gt, is it still possible to sample in the original gt，So why does its size affect the final result？