jiwei0921 / dmra Goto Github PK

I found the depth size is 256x256 in DUT-RGBD-400*600, can you provide the original depth file (400x600)?

i want to train on my own dataset, and i found the format of images in masks is 1-bit .png, i am wondering how to make 1-bit .png?

In your code you use "depth = np.array(depth, dtype=np.uint8)" to load your depth image, I was wondering maybe this is not okay. Because of the depth value is much bigger than 255. If you use 'np.uint8', the depth value is not correct when loaded.
Thank you very much. @jiwei0921

Thanks for your code and paper!
I notice that in your implementation, the ConvLSTMCELL returns the o as the output of the cell.
x, new_c, new_o = getattr(self, name)(x, h, c) # ConvLSTMCell forward
(line 316 in fusion.py)
However, i notice that in the pytorch implementation, the convlstm use the h as the output the ConvLSTMCELL. Could you please tell me the difference and the reason?

RuntimeError: The size of tensor a (150) must match the size of tensor b (152) at non-singleton dimension 3
how to solve it?

Hi Jiwei!

Why you try a little bit more bigger learning rate in your training phase?

Or

You try some bigger lr like lr = 3e-4 or lr = 1e-6, can you suggest some useful experiential value?

when I run demo.py, I find that drb5's shape = (1,64,100,152), but drb1,drb2... their shape = (1,64,100,150)

Hello,

Thank you for your nice work.
I have some questions with the Table 1 in your paper.
Did you train/test RGB methods with the RGB images in the RGBD datasets from scratch? or finetune their pretrained model?
For PiCANet, which backbone did you use for your table? Is it VGG16 or ResNet?

Thank you.
Best regards,
Ahyun

Please answer this question during the test. Thank you