nv-tlabs / dib-r Goto Github PK

Hi,
Is it possible to provide access to your pretrained model and let us know the format of the test list so we can run the test_all script with some test data? Thanks.

Dear All,

2 points,

a) there is a typo in the README cmd, should be pip install -r requirements.txt.
b) pip fails at torch
ERROR: Could not find a version that satisfies the requirement torch==1.1.0 (from -r requirements.txt (line 3)) (from versions: 0.1.2, 0.1.2.post1, 0.1.2.post2)
ERROR: No matching distribution found for torch==1.1.0 (from -r requirements.txt (line 3))

PIP version:
pip 20.0.1 from c:\users\me\appdata\local\programs\python\python37-32\lib\site-packages\pip (python 3.7)

I have Python 3.7, Cuda 10.0.130, Visual Studio 2017, Windows 10, and am using Pip.

I'm on the python build.py install step which is giving me this error:

1 error detected in the compilation of "C:/Users/me/AppData/Local/Temp/tmpxft_000047dc_00000000-8_dr_cuda_back.cpp4.ii". error: command 'C:\\Program Files\\NVIDIA GPU Computing Toolkit\\CUDA\\v10.0\\bin\\nvcc.exe' failed with exit status 2

Can I please receive advice on how to fix this?

I'm currently rendering using RenderMeshColor as the rendering class. If I render 2 meshes, the nearer mesh does not occlude the further away mesh. I realise now that this is causing the issue in #8

Here is an example:

How can I render multiple meshes with vertex colour, with the correct occlusion?
Thanks!

Hi,

Thanks for the excellent paper!

Just wondering where is the example model for inference?

Also, is there a template that lists the format of "test_list.txt"?

Thanks,

For my application, I am rendering using the per-vertex texture mode. I need to compute which texture values are visible in the rendering. What is the best way to do this? I thought there might be a way to exploit the texture gradients, since I would just need to know which texture values are assigned a non-zero gradient after rendering.

At the moment, how I do it is to render UV values of the vertices, project the vertices to image space, and consider a vertex to be visible if its UV value is close to the UV value of where it is projected in the rendered UV. This fails close to the edge of the mesh because the background has a rendered UV of 0, and a lot of vertices close to the edge get excluded, even though they are visible.

Hello,

I am currently trying to train a model with the ShapeNetP2M Dataset and the corresponding test list (I read in another issue that you used this data yourself).
The first issue is that the received DataProvider

data = get_data_loaders(filelist, imsz, viewnum, mode='test', bs=BATCH_SIZE, numworkers=numworkers, data_folder=data_folder)

has no Data (Data is none).

After I tried to overcome this issue by simply avoiding the torch.dataloader script and use your prepare_instance in dataloader_multiview_blender, I stumbled across this error:

imnp = da['view%d' % j]['im'] TypeError: 'int' object is not subscriptable

I think it is because the load_in_cam function needs .npy files which are not provided in the ShapeNetP2M Dataset. There are only .dat files besides the png images
What is the correct way to load this dataset? Am I missing something? Thanks in advance!

Looks like there's some files/folders missing in my directory? I am not sure, I'm running on windows 10 and don't know how to format the test command for windows. When I run on bash it seems to not recognize torch.

DIB-R\dib-render>python test-all.py --g_model_dir ./checkpoints/g_model.pth --svfolder ./prediction --data_folder ./dataset --filelist ./test_list.txt

| Configurations |

IMG_DIM: 64
N_CHANNELS: 4
BATCH_SIZE: 64
FILELIST: ./test_list.txt
TOTAL_EPOCH: 1000
ITERS_PER_LOG: 50
VERBOSE: True

Building dataloaders
./dataset
Traceback (most recent call last):
File "test-all.py", line 108, in
bs=BATCH_SIZE, numworkers=numworkers,data_folder=data_folder)
File "../utils\dataloader\dataloader_multiview_blender.py", line 167, in get_data_loaders
mode=mode, datadebug=False, classes=classes, data_folder=data_folder)
File "../utils\dataloader\dataloader_multiview_blender.py", line 53, in init
with open(file_list, 'r') as f:
FileNotFoundError: [Errno 2] No such file or directory: './test_list.txt'

Is it possible to render faces with the texture on both sides of the face, ie both orientations?

Hi,

Many thanks for your work.!

I use your lib to render a image, but it seems to be slow. I find many loops in your code. Can it speed up ? Or any suggestion?

Best,

I have collected the images of some gadgets from internet, and now I want to train DIB-R on those images. How do I do that?

at python build.py install getting the error:

C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\include\yvals.h(12): fatal error C1083: Cannot open include file: 'crtdbg.h': No such file or directory
error: command 'C:\Program Files (x86)\Microsoft Visual Studio\2019\Community\VC\Tools\MSVC\14.27.29110\bin\HostX86\x64\cl.exe' failed with exit status 2

Possibly an error with versions, currently running python 3.8.3 (64bit) and CUDA 11.2.0.

What versions of Python, Visual Studio and CUDA are best suited to run this?

Hi there, thanks for releasing the code for this awesome project! Quick question regarding the evaluation you used:

when I compare the 3D IoU results in your paper with the ones reported by previous SOTA DR algorithms, they are a bit different (e.g. SoftRas reported 62% whereas it is 59% in your paper). Do you know where the differences come from? I suppose you obtained such results by running their algo with your evaluation but it seems quite similar to me

Thanks

Could you plz sooner share your code? This is really an amazing progress, I'm really hungry for it.

Thank you very much!!!!!!

In my use case, I am getting a bug where the furthest face from the camera is rendered, rather than the closest. So when I render a car on a turntable, I see the interior of the car. This is not due to the orientation of the faces, since I have tried rendering the faces from both orientations as mentioned in #4
Do you have any idea what may be causing this? Thanks!

How can I render single images from the CUB dataset and visualize it in 3D as described in the paper? thanks

I'm trying to get this project up and running on my own Windows machine. I installed the dependencies, but was unable compile the DIB-Render as shown in the README because I have an Intel-based GPU (unsupported by NVIDIA's CUDA).

I thought to myself that maybe I can simply run the project on a CPU, and I saw in the test-all.py file that there seems to be some support for CPU-based machines in this project. From line 130:

device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')

However, when I tried to build the cuda_dib_render I realized that I must have CUDA on my machine. So it's not clear from the docs - do I have to have a CUDA-enabled GPU, or can I run this project on a normal Intel GPU? Or even a CPU?

If I have to run it over a CUDA-enabled GPU, do you have a recommended setup for a compatible machine on Google Cloud Platform? Would help a lot to future developers who want to test out this nice project by avoiding these early pitfalls.

Thanks!

Hey, thanks so much for this! For my use case I would like to render using a texture with more than 3 channels. It is not currently possible:
"colors_bxfx3d must be same point size (dr_forward_batch at dr.cpp:90)"

Could you please make this feature available as an option?

Hi, nice work! Do you use the CUB-200-2011 dataset provided by CMR? Because it is different from the CUB dataset provided by the official. I found that the results presented in your paper are different from the result I evaluated on the data provided by CMR. btw, I used the model trained 500 epochs. And the test set contains 2874 images. If not, could you provide the dataset you used to train and evaluate your model?

Hello,

I am a newbie in both python and dlib software. Is there a simple document for dummies on how
to use this software. I could not make out from the description provided on the website.

Thanks.

Now it only supports color and mask, how can I get depth?