search

TruongKhang / cds-mvsnet

[ICLR2022] Curvature-guided dynamic scale networks for Multi-view Stereo
117 stars 7 forks source link

Problem with demo #17

Closed denred0 closed 1 year ago

denred0 commented 1 year ago

I tried this demo and got strange results. I set the environment and installed all libs. Start the demo with the comand: python demo.py --imagedir=data/nmail6 --calib=calib/nmail3.txt --stride=2 --buffer 384 --mvsnet_ckpt cds_mvsnet.pth --disable_vis After frame 23 processing stops. I found out that this line is the source of the error: DROID-SLAM\droid_slam\geom\projective_ops.py

# transform
Gij = poses[:,jj] * poses[:,ii].inv()

Since the determinant of the matrix poses is equal to zero, the operation inv() cannot be performed. It looks like poses not updated.

Do you know where is can be problem? Thank you.

TruongKhang commented 1 year ago

Hi @denred0, it looks like from the DROID-SLAM code. Unfortunately, I didn't dig into this code in detail so I'm not sure. But I ran the demo and there was no problem at all. Can you share a screenshot of your errors came out? And how much is your GPU memory?

denred0 commented 1 year ago

Hi @TruongKhang, thanks for quick response. I got error Process finished with exit code -1073741819 (0xC0000005). I have only 4Gb of GPU on my laptop and think this error can be connected with small amount of GPU memory. Today I want to test code on Colab to prove my theory about small amount of GPU memory.

TruongKhang commented 1 year ago

Oh, I see. I think you should use a GPU of at least 8 GB. Or, you can try to reduce the image_size to test on your current GPU.

denred0 commented 1 year ago

I tried to reduce the image_size but got error: The expanded size of the tensor (224) must match the existing size (448) at non-singleton dimension 2. Target sizes: [3, 128, 224]. Tensor sizes: [3, 256, 448]

Probably it is necessary to somehow change the code to make it work.

TruongKhang commented 1 year ago

humm, I see. So please let me know the result when you try on Google Colab! :D

denred0 commented 1 year ago

As a result, I was able to run the demo in the colab. The code consumes 15Gb RAM during installation and 5.5Gb GPU during inference. Thanks for your help.