hjwdzh / adversarialtexture Goto Github PK

I'm trying to use another method to render your dataset, but the coordinate seems wired.

All data is downloaded from your README file.

And it has already been compiled and passed.

Sens.Parse(c_char_p(filename.encode('utf8')))

Process finished with exit code 134

It's a great work and thank you for sharing the code!

I have tried to run the code with your chair data, and it did show a significant improvement compared to L1 result.

Then, I edit render_scan.py file to read data from Zhou's fountain model (I choose around 30 RGBD frames as key frames, and use code from "Let there be color" to get the obj and mtl files from all 30 key frames.), I run the code with default parameters (λ=10.0, iter=4001).

But I don't get the fine result as the supplemental shows.

Here's what I get:

Before (L1):

After Iterations:

Is anything missing to run such RGBD datasets?

Thanks again for the interesting work!

Hi, thank you for sharing source code, it's a fancy work!

I have a question about using custom dasaset. Since I have some RGBD dataset, I composed it to .sens file like BundleFusion (e.g. apt0.sens). But it couldn't work well using render_scan.py.

Any suggestion or is there a detail about how your .sens file composed?

Thank you~

Hello, I tried to download the data with commands in the readme at data directory, and they all response 404.
Is there any thing wrong with the data source?
Expect for your reply!

OSError: ./Rasterizer/libSens.so: undefined symbol: _ZN2cv8imdecodeERKNS_11_InputArrayEi

Is there any suggestion?

thanks

Hello,could you write the Dependencies in readme?eg. I install TF2 but the code need TF1.I would be very grateful if you could accept my opinion.Thanks.

Dear Jingwei, I have other two questions about the code

• In Line128 of dataset.py, why 1-uv?

• In Line 146 of model.py, why discrim_loss is set to 0 when it is greater than gen_loss_GAN?

Very impressive work! While checking the dataset file, I found you utilize the cache to ensure no repeat calculation for each source image. But it will lead to one-to-one matching during training, rather than selecting the target view randomly. Is it right?