About APs,APm,APL

[LEE-JAEHA]

Hello. I'm looking at your code and the performance you posted on Readme is as below.

Average Precision (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.207 Average Precision (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.410 Average Precision (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.495

But when I actually turned it around, the performance isn't that good, is the parameter setting different from the one posted on GitHub?

And what kind of backbone network is the performance that comes from?

And one more could you tell me the difference from the official fcos code?

Thank you.

[zhenghao977]

@LEE-JAEHA Hello, can you provide more training details?

[LEE-JAEHA]

@zhenghao977

Thank you for your answer. In the case of training detail, I did exactly what you wrote on config.py, and used the backbone that is pre-trained resnet101 on ImageNet

In the author's paper uses resnet 50 and the Aps Aptm APl came out with 19.7 39.7 48.8, but your code performance was not get that result.

Can I know your training detail?

---

class DefaultConfig():

backbone

pretrained=True
freeze_stage_1=True
freeze_bn=True

#fpn
fpn_out_channels=256
use_p5=True

#head
class_num=80
use_GN_head=True
prior=0.01
add_centerness=True
cnt_on_reg=True

#training
strides=[8,16,32,64,128]
limit_range=[[-1,64],[64,128],[128,256],[256,512],[512,999999]]

#inference
score_threshold=0.05
nms_iou_threshold=0.6
max_detection_boxes_num=1000

[zhenghao977]

@LEE-JAEHA OK, I see. I trained coco2017 on 4 Tesla-V100 .batch_size is 16, init lr=1e-2 , total epoch is 24. Other settings are same as config.py.

[LEE-JAEHA]

@zhenghao977
Could yo tell me the backbone network?

[zhenghao977]

@LEE-JAEHA All models are resnet50

[LEE-JAEHA]

@zhenghao977 and could you tell me why you set the img size with [800,1333]?? I don't get that much performance even if I train, is there any effect on the image size?

Thank you so much

[zhenghao977]

@LEE-JAEHA img size has a big effect, if you use low resolution, the results must be worse. You can refer to the original paper for this resolution(800,1333).