The performance is very bad and disappointed when using fcos

[alexwq100]

hi, @tianzhi0549 I tried your approach , I find that the performance of this method is very inferior compared with maskrcnn benchmark based on resnet 50 backbone. From the comparison, the map of maskrcnn can be 75%+, but your is only 40%, and expecially for the hight recall , the precision is very low. I used the code in mmdet. And in mmdet, I try the default configuration in https://github.com/open-mmlab/mmdetection/blob/master/configs/fcos/fcos_r50_caffe_fpn_gn_1x_4gpu.py ,the result is very bad. For just comparison, I change the default configuration into the following as the configuration in maskrcnn: In the fcos default configuration ,the output of stage 2 in resnet50 is not used.Because, In my dataset ,there are some tiny object and they may be put very closely. So I think the low feature map should be used. But the result still very bad compared with maskrcnn. Do you have any idea?

[tianzhi0549]

@alexwq100 discussion #31 might be helpful to you. Moreover, I think you should do an error analysis of your detection results. It is difficult for me to help you without these details.

[alexwq100]

thanks for your quick reply! I am doing the analysis. but I think, for the same dataset, I use the maskrcnn default configuration and your default configuration,the performance difference is huge. I think this is abnormal By the way, how to set this parameter https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/anchor_heads/fcos_head.py#L22? should I decide it according to the feature map fov ?

[tianzhi0549]

@alexwq100 I think it somehow depends on the size distribution of the objects in your dataset. Also, it is related to the receptive field of the feature maps.

[alexwq100]

you said "I think it somehow depends on the size distribution of the objects in your dataset. ",,could you explain it in more detail? At the beginning I want to use it to replace the standard rpn in the maskrcnn to enhance the overall recall, do you think this is reasonable?

[tianzhi0549]

@alexwq100 I suggest that you slightly tune the sizes to see what happens. It is difficult to give suggestions due to lack of details of your dataset.

[alexwq100]

@tianzhi0549 thanks for your reply, "the size" you mentioned is what? you mean the regress_ranges=((-1, 64), (64, 128), (128, 256), (256, 512), (512, INF)) parameters? if so, I am doing experiment now

[tianzhi0549]

@alexwq100 yes.

[alexwq100]

@tianzhi0549 hi，sorry for asking again. In the experiment of part 5 in your paper, you mention "Compared to RPNs with FPN [11], we replace anchor boxes with the method in FCOS. Moreover, we add GN into the layers in FPN heads, which can make our training more stable.". I try to change the configuration for self.lateral_convs and self.fpn_convs in fpn from norm_cfg =None to norm_cfg=dict(type='GN', num_groups=32, requires_grad=True) in the mmdet framwork, see https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/necks/fpn.py#L59 and https://github.com/open-mmlab/mmdetection/blob/master/mmdet/models/necks/fpn.py#L68. And my configuration py file is as: But the result is not get improved as your paper reports. Could you please help me check the modification is right ? thks a lot

[tianzhi0549]

@alexwq100 Can you post your results here?

[alexwq100]

fcos result

retina result

maskrcnn result

all use the default configuration, the fcos is very disappointed compared with other two.

[alexwq100]

I try you said ,but the result is not get improved, and I also replace the rpn in maskrcnn with fcos, the result get improved, but not yet outperform the result given by maskrcnn

[JerryIndus]

fcos result

retina result

maskrcnn result

all use the default configuration, the fcos is very disappointed compared with other two.

Excuse me, I met the same problem and I want to ask whether you have resolved it or not