With the modified DLA-34 backbone, why only IDAUP the first three layers and not all four

[liu-mengyang]

In the default configuration, 'first_level' is 2 and 'last_level' is 5.
In line 443 of pose_dla_dcn.py,👇

self.ida_up = IDAUp(out_channel, channels[self.first_level:self.last_level],
             [2 ** i for i in range(self.last_level - self.first_level)])

This caused 'dla_up' to generate only three up_f and coderange(self.last_level - self.first_level) also appear in IDAUP operation.

I think this conflicts with the description in the paper, as shown below:

Can someone tell me if this is my fault or just a special design?🥺

[hhaAndroid]

I have the same question

[janosfoeth]

I have the same issue and I'd say this is an implementational flaw.
@liu-mengyang Did you fix the implementation so that the code is coherent with the description in the paper?

Do I just need to add one to the range of IDA_Up layers?

So in detail - change this:

CenterNet/src/lib/models/networks/pose_dla_dcn.py

Lines 442 to 443 in 2b7692c

	self.ida_up = IDAUp(out_channel, channels[self.first_level:self.last_level],
	[2 ** i for i in range(self.last_level - self.first_level)])

to this:

self.ida_up = IDAUp(out_channel, channels[self.first_level:self.last_level + 1], 
		[2 ** i for i in range(self.last_level - self.first_level + 1)])

and this:

CenterNet/src/lib/models/networks/pose_dla_dcn.py

Line 475 in 2b7692c

for i in range(self.last_level - self.first_level):

to this:
for i in range(self.last_level - self.first_level + 1):

[NosremeC]

I realized this problem long time ago. I think the reason behind is simply that the 4th layer does not perform well.

I have tested this model on waymo dataset. For waymo adding the 4th layer resulted in slower converging and worse performamce. I think it is possibly due to the small number of downsample operations in this network. If you compare resnet with dla, you'll find resnet is deeper and thus its deeper layers actually contain useful information. For dla the network is still realatively shallow even at the 4th layer. I guss adding the 4th layers will possibly introuce more noise than information.