Maybe something inappropriate in limiting the range of tx or ty ? about racnn-pytorch

[FanvelChen]

I can got the motivation of your manual change, but what I mean is the direction of comparison maybe is incorrect.
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) < tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) < ty else (in_size/3*2)
in my understanding it should be
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) > tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) > ty else (in_size/3*2)

Another part is in line 101 in ./models/RACNN.py
xatt_cropped = xatt[:, h_off : h_end, w_off : w_end]
should be
xatt_cropped = xatt[:, w_off : w_end, h_off : h_end]
because your x is vertically increasing, and your y is horizontally increasing. According to the corresponding index relation, I think the slice should be changed like this.

[zjuPeco]

I can got the motivation of your manual change, but what I mean is the direction of comparison maybe is incorrect.
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) < tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) < ty else (in_size/3*2)
in my understanding it should be
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) > tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) > ty else (in_size/3*2)

Another part is in line 101 in ./models/RACNN.py
xatt_cropped = xatt[:, h_off : h_end, w_off : w_end]
should be
xatt_cropped = xatt[:, w_off : w_end, h_off : h_end]
because your x is vertically increasing, and your y is horizontally increasing. According to the corresponding index relation, I think the slice should be changed like this.

the index order is (C, H, W).
change line 78-79 in ./models/RACNN.py

x = torch.stack([unit.t()] * 3)
y = torch.stack([unit] * 3)

to

x = torch.stack([unit] * 3) 
y = torch.stack([unit.t()] * 3)

works for me.
because
mk = (h(x-w_off) - h(x-w_end)) * (h(y-h_off) - h(y-h_end))

[jeong-tae]

i found that if cropped image size is too small performance is dropped. so i manually change the cropping rule bcuz i do not have pretrain APN.

For optimization, see “The unreasonable effec- tiveness of noisy data for fine-grained recognition” this paper. i also dont know about detail well. In the RACNN paper, their intention is to keep a consistent direction of optimization.

[jeong-tae]

oh i see. i misunderstand your first question. i will check inquality direction.

second one is, as far as i know, the index order maybe startwith height and then width so that slice was set as 101 line. i will check again

[libing125]

I can got the motivation of your manual change, but what I mean is the direction of comparison maybe is incorrect.
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) < tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) < ty else (in_size/3*2)
in my understanding it should be
tx = tx if tx > (in_size/3) else in_size/3
tx = tx if (in_size/3*2) > tx else (in_size/3*2)
ty = ty if ty > (in_size/3) else in_size/3
ty = ty if (in_size/3*2) > ty else (in_size/3*2)
Another part is in line 101 in ./models/RACNN.py
xatt_cropped = xatt[:, h_off : h_end, w_off : w_end]
should be
xatt_cropped = xatt[:, w_off : w_end, h_off : h_end]
because your x is vertically increasing, and your y is horizontally increasing. According to the corresponding index relation, I think the slice should be changed like this.

the index order is (C, H, W).
change line 78-79 in ./models/RACNN.py

x = torch.stack([unit.t()] * 3)
y = torch.stack([unit] * 3)

to

x = torch.stack([unit] * 3) 
y = torch.stack([unit.t()] * 3)

works for me.
because
mk = (h(x-w_off) - h(x-w_end)) * (h(y-h_off) - h(y-h_end))

I think you are right. How is your result? Can you reproduce the result of the paper?

[libing125]

RACNN-pytorch/models/RACNN.py

Line 121 in 35716d5

x = torch.stack([torch.arange(0, in_size)] * in_size).t()

Line 121 should also be changed. Remove the ".t()".

[YananGu]

@libing125 Hi, did you reproduce the result of the paper by this code?