Question about center-ness branch in FCOS

[WZMIAOMIAO]

Hi, thank you for your great work. I'm learning FCOS these days. I find some differences about position of center-ness between code and paper. In paper(https://arxiv.org/abs/1904.01355), the center-ness branch is put together with the classification branch.

But in the code, the center-ness and regression branches are put together.

vision/torchvision/models/detection/fcos.py

Lines 202 to 233 in a1232c2

	class FCOSRegressionHead(nn.Module):
	"""
	A regression head for use in FCOS.
	Args:
	in_channels (int): number of channels of the input feature
	num_anchors (int): number of anchors to be predicted
	num_convs (Optional[int]): number of conv layer. Default: 4.
	norm_layer: Module specifying the normalization layer to use.
	"""
	def __init__(
	self,
	in_channels: int,
	num_anchors: int,
	num_convs: int = 4,
	norm_layer: Optional[Callable[..., nn.Module]] = None,
	):
	super().__init__()
	if norm_layer is None:
	norm_layer = partial(nn.GroupNorm, 32)
	conv = []
	for _ in range(num_convs):
	conv.append(nn.Conv2d(in_channels, in_channels, kernel_size=3, stride=1, padding=1))
	conv.append(norm_layer(in_channels))
	conv.append(nn.ReLU())
	self.conv = nn.Sequential(*conv)
	self.bbox_reg = nn.Conv2d(in_channels, num_anchors * 4, kernel_size=3, stride=1, padding=1)
	self.bbox_ctrness = nn.Conv2d(in_channels, num_anchors * 1, kernel_size=3, stride=1, padding=1)

Could you tell me why? thanks.

[datumbox]

@zhiqwang @xiaohu2015 @jdsgomes Could you give a reply on this since you know the technique inside-out?

[xiaohu2015]

@WZMIAOMIAO Hi, you can read the updated paper (https://arxiv.org/pdf/2006.09214.pdf, page6 note). The center-ness branch added on regression branch can get better result.

This is different from our conference version which positions the
center-ness on the classification branch, but it has been shown that
positioning it on the regression branch can obtain better performance.

[WZMIAOMIAO]

@WZMIAOMIAO Hi, you can read the updated paper (https://arxiv.org/pdf/2006.09214.pdf, page6 note). The center-ness branch added on regression branch can get better result.

This is different from our conference version which positions the
center-ness on the classification branch, but it has been shown that
positioning it on the regression branch can obtain better performance.

Wow, that is great, Thank you.

[WZMIAOMIAO]

It is suggested to update the paper link in the description document.

vision/torchvision/models/detection/fcos.py

Line 682 in a1232c2

Reference: `FCOS: Fully Convolutional One-Stage Object Detection <https://arxiv.org/abs/1904.01355>`_.

[datumbox]

Yes that might be a good idea, thanks for flagging. We should probably highlight both papers. You want to bring a PR?

[WZMIAOMIAO]

This is my honor.

[datumbox]

@WZMIAOMIAO No, definitely ours for helping us improve the documentation.

Reopening the ticket until the PR lands. Feel free to tag me on the PR for reviews. Thanks a lot for the help!