search

xingyizhou / CenterNet2

Two-stage CenterNet
Apache License 2.0
1.2k stars 189 forks source link

how to use centernet* only #35

Open curry8 opened 3 years ago

curry8 commented 3 years ago

hi,xingyi, thanks for your work, I I just want to use centernet*(FPN) to do pose estimate instead of two stage detector , what should I do, looking for your reply, thanks again!

xingyizhou commented 3 years ago

Hi, thank you for your interest. Unfortunately, we only have the object detection part in this repo and don't have one-stage pose estimation :(

Lg955 commented 3 years ago

Hi, thank you for your interest. Unfortunately, we only have the object detection part in this repo and don't have one-stage pose estimation :(

Hi, if I have 10000 images and 5 classes, what is the suitable MAX_ITER for training Centernet2?

curry8 commented 3 years ago

thank you for your reply! I have one more question about FPN, you assign positive and negative samples by target size to different stride in FPN. I wonder if the same target will be assigned to both positive and negative in different stride fpn feature map Won’t it affect the training,looking for your reply.

xingyizhou commented 3 years ago

Hi @curry8 , great questions. Yes, the same target will be assigned to positive in some FPN levels (Usually one level, but can be more than one). For other levels, we ignore the negative losses in the center location. However, just assigning them is also fine, it will be just within 1 mAP worse.

curry8 commented 3 years ago

Hi @xingyizhou xingyi, I train model with FPN after a longtime of training, I found that model with fpn performance is basically the same as model without FPN I set the area of less than 50*50 as a small target by the width and height of the target in the original img medium and large target, assigned it to a feature map stride 4 8 16 Do you think the test results are normal? You uesd fpn in centernet2. Could the fpn bring improvement looking forward to your reply thanks!