Centernet: Why resnet backbones do not perform well on CenterNet?

[Epiphqny]

The AP is even less than the much smaller DLA-34, do you have any idea?

[lbin]

which backbones do u use? Res50 or Res34?

[Epiphqny]

which backbones do u use? Res50 or Res34?

As shown in the README of this project, the resnet-101 is about 3 points lower than dla-34 backbone : DLA-34(37.4), Resnet-101(34.6)

[lbin]

which backbones do u use? Res50 or Res34?

As shown in the README of this project, the resnet-101 is about 3 points lower than dla-34 backbone : DLA-34(37.4), Resnet-101(34.6)

May be DLA has more connections? This is strange and instesting, because on Imagenet, DLA's top1 accuracy is much lower than res50 or res101.

[Epiphqny]

which backbones do u use? Res50 or Res34?

As shown in the README of this project, the resnet-101 is about 3 points lower than dla-34 backbone : DLA-34(37.4), Resnet-101(34.6)

May be DLA has more connections? This is strange and instesting, because on Imagenet, DLA's top1 accuracy is much lower than res50 or res101.

Yes, that's my doubt, i want to ask how much degree the CenterNet relies on the backbone.

[xingyizhou]

This is a very good question. We need to distinguish two concepts: "backbone" and "network". "network" is backbone + upsampling layers (or "neck" in some other papers). In our resnet_dcn models or msra_resnet models, the upsampling layers are light (3 upconv layers), while in our dla models, the upsampling layers are large (DLA_up + IDA_up, please refer to the code and fig.6 in the supplementary). I would say the upsampling layers matter a lot, and they are not easily transferable across backbones: the default DLA_up requires keeping the original channels in the backbone (for DLA, the #channels from 4x stride to 32x stride are 64, 128, 256, 512, for ResNets, they are 256, 512, 1024, 2048), and using IDA_up for ResNets will be very expensive. I can only afford to try Res18 + DLA_up, the performance is much better than Res18+dcn upconvs and close to DLA34+IDA_up on pascal. I haven't tried DLA + dcn upconvs.

[Epiphqny]

This is a very good question. We need to distinguish two concepts: "backbone" and "network". "network" is backbone + upsampling layers (or "neck" in some other papers). In our resnet_dcn models or msra_resnet models, the upsampling layers are light (3 upconv layers), while in our dla models, the upsampling layers are large (DLA_up + IDA_up, please refer to the code and fig.6 in the supplementary). I would say the upsampling layers matter a lot, and they are not easily transferable across backbones: the default DLA_up requires keeping the original channels in the backbone (for DLA, the #channels from 4x stride to 32x stride are 64, 128, 256, 512, for ResNets, they are 256, 512, 1024, 2048), and using IDA_up for ResNets will be very expensive. I can only afford to try Res18 + DLA_up, the performance is much better than Res18+dcn upconvs and close to DLA34+IDA_up on pascal. I haven't tried DLA + dcn upconvs.

Thanks for your detailed explanation. Can i understand it as that the keypoint estimation is more like semantic segmentation, which requires large resolution and dense prediction, therefore upsampling is critical to it?

[niaoyu]

This is a very good question. We need to distinguish two concepts: "backbone" and "network". "network" is backbone + upsampling layers (or "neck" in some other papers). In our resnet_dcn models or msra_resnet models, the upsampling layers are light (3 upconv layers), while in our dla models, the upsampling layers are large (DLA_up + IDA_up, please refer to the code and fig.6 in the supplementary). I would say the upsampling layers matter a lot, and they are not easily transferable across backbones: the default DLA_up requires keeping the original channels in the backbone (for DLA, the #channels from 4x stride to 32x stride are 64, 128, 256, 512, for ResNets, they are 256, 512, 1024, 2048), and using IDA_up for ResNets will be very expensive. I can only afford to try Res18 + DLA_up, the performance is much better than Res18+dcn upconvs and close to DLA34+IDA_up on pascal. I haven't tried DLA + dcn upconvs.

Thanks for your patience.

1. You just mentioned the two structure "IDA_up" is more expensive than "DLA_up" so you choose to use "res + dla_up", but "DLA_up" is consturcted with some "IDA_up".
   https://github.com/ucbdrive/dla/issues/14#issuecomment-524699388
   Why IDA_up consumes more?
2. Just as you said, the output dimensions of "DLA_up" keep constant and the "IDA_up" will change all layers to the same output dimensions(upsample). So in your "res + dla_up", is it still need more layers after "DLA_up" and before "head layers"?Or is it just a mistake which should be " res + IDA_up"?

1. I tried to make some improvements about resnet.
   3.1) resnet + fpn +dcn+ upsample + add
   3.2) resnet +ida_up （which actually is .. dcn+ upsample + add+dcn）
   Which one would be better?

---

不好意思，英文可能不太清晰。我好奇的是，

1. 这里提到的"IDA_up"资源消耗比"DLAUp"大，但实际上"DLAUp"是由多个"IDA_up"组成的.
   https://github.com/ucbdrive/dla/issues/14#issuecomment-524699388
   为什么反而IDA_up资源消耗大呢。
2. "DLA_up"模块的各个层的输出纬度和输入一致保持不变的，在原始的版本中是需要"IDA_up"将所有层统一改为相关的维度-上采样-才能进行add操作。那后续提到的“res+dla_up”中，如果没有ida_up，dla_up后面连接的是什么结构，才能得到维度相同的featuremap（后续连上head）。或者说，写错了实际上使用的是“res+ida_up”？

1. 这里想改进resnet效果。
   本想使用res+fpn，然后通过的upsample+add+dcn得到feature map。
   对比下，感觉可以采用res5+ida_up，就可以得到比较好的featuremap。
   哪种更合理些，有建议吗

[sisrfeng]

They are several variants outperforming ResNet. Why do some researchers still use the old ResNet rather than ResNext or SENet?
Is it because ResNet is GPU-friendly?
如何评价ResNeSt：Split-Attention Networks？ - 知乎
https://www.zhihu.com/question/388637660/answer/1162087825

[menggui1993]

This is a very good question. We need to distinguish two concepts: "backbone" and "network". "network" is backbone + upsampling layers (or "neck" in some other papers). In our resnet_dcn models or msra_resnet models, the upsampling layers are light (3 upconv layers), while in our dla models, the upsampling layers are large (DLA_up + IDA_up, please refer to the code and fig.6 in the supplementary). I would say the upsampling layers matter a lot, and they are not easily transferable across backbones: the default DLA_up requires keeping the original channels in the backbone (for DLA, the #channels from 4x stride to 32x stride are 64, 128, 256, 512, for ResNets, they are 256, 512, 1024, 2048), and using IDA_up for ResNets will be very expensive. I can only afford to try Res18 + DLA_up, the performance is much better than Res18+dcn upconvs and close to DLA34+IDA_up on pascal. I haven't tried DLA + dcn upconvs.

Thanks for your patience.

1. You just mentioned the two structure "IDA_up" is more expensive than "DLA_up" so you choose to use "res + dla_up", but "DLA_up" is consturcted with some "IDA_up".
   ucbdrive/dla#14 (comment)
   Why IDA_up consumes more?
2. Just as you said, the output dimensions of "DLA_up" keep constant and the "IDA_up" will change all layers to the same output dimensions(upsample). So in your "res + dla_up", is it still need more layers after "DLA_up" and before "head layers"?Or is it just a mistake which should be " res + IDA_up"?
3. I tried to make some improvements about resnet.
   3.1) resnet + fpn +dcn+ upsample + add
   3.2) resnet +ida_up （which actually is .. dcn+ upsample + add+dcn）
   Which one would be better?

不好意思，英文可能不太清晰。我好奇的是，

1. 这里提到的"IDA_up"资源消耗比"DLAUp"大，但实际上"DLAUp"是由多个"IDA_up"组成的.
   ucbdrive/dla#14 (comment)
   为什么反而IDA_up资源消耗大呢。
2. "DLA_up"模块的各个层的输出纬度和输入一致保持不变的，在原始的版本中是需要"IDA_up"将所有层统一改为相关的维度-上采样-才能进行add操作。那后续提到的“res+dla_up”中，如果没有ida_up，dla_up后面连接的是什么结构，才能得到维度相同的featuremap（后续连上head）。或者说，写错了实际上使用的是“res+ida_up”？
3. 这里想改进resnet效果。
   本想使用res+fpn，然后通过的upsample+add+dcn得到feature map。
   对比下，感觉可以采用res5+ida_up，就可以得到比较好的featuremap。
   哪种更合理些，有建议吗

@niaoyu Hi, have you tried with res+fpn?