Matthias Niessner

Пікірлер

@floribertjackalope2606 3 күн бұрын

is it intended that the course site was taken down?

@onurcanisler 6 күн бұрын

Oh dear god. Surely the hardest lecture of all.

@mrburns366 23 күн бұрын

So.. LA Noire 2? 😁

@AaliDGr8 23 күн бұрын

how to use it stpd act WTH did not you give us working link ??? plz

@weelianglien687 Ай бұрын

i wonder in the AlexNet (e.g. 1st conv layer) example, should the kernels be labelled as 11x11x3 due to the RGB, unless the usage of blue colour layers in all the stages signifies that this is only an illustration for the B layer?

@ONDANOTA Ай бұрын

thanks! very informative

@alex23361 Ай бұрын

fantastic

@TokamakPower Ай бұрын

My neurons do tend to dropout during tests.

@shishen5253 Ай бұрын

Very impressive work！Will the code for this paper be open source?

@AR-vb4xy Ай бұрын

Very interesting video but I suggest a improvement with respect to the display of the lecture: The block on the lower right where the recording of the professor is, should be very small or removed alltogether.

@interfect Ай бұрын

Thank you for the great lectures! Is it somehow possible to get the slides for parts 8-12?

@dddd-wf6fn Ай бұрын

如果有很多万级带有语义的三维模型，是否可作为训练数据，用three.js加载后，自动输出训练数据？

@MisterWealth 2 ай бұрын

When will the code be made available?

@shan_420 2 ай бұрын

24:00 I think it's a bit confusing with f=Wx and in the image it's xW=f, specially when talking about the dimensionality of W.

@bakikucukcakiroglu 2 ай бұрын

using the test data over and over again makes it the second phase validation data so doing that can be considered as skipping the test phase.

@rallyworld3417 3 ай бұрын

Impressive

@kasemir0 3 ай бұрын

body, how can i use this code on the Colab? please, help me. tks

@M_a_t_z_ee 3 ай бұрын

Great introductory lecture. I'm excited about the following ones as well as the programming exercises. 😀

@yimloo60 3 ай бұрын

Thank you! Deep Learning! Thats my first time to recognize that i have a goddamn brain! :)

@eskimo2616 3 ай бұрын

19:11 what does "clamp it to zero" mean?

@ulascanzorer 3 ай бұрын

I think it just means we set zero as our minimum so that the values can't go lower than that.

@M_a_t_z_ee 3 ай бұрын

It means that you take two argments for the maximum function: 0 and the function based on previous inputs. This translates to the ReLU (rectified linear unit) activation function. If the second argument is bigger than 0, the max function evaluates to that argument. If the second arguments is negative, the max function evaluates to 0. It "clamping" all negative outputs from the previous layer to 0.

@DaveDFX 4 ай бұрын

Amazing ! Game changer for avatars

@PakkaponPhongtawee 4 ай бұрын

Could you please upload the supplementary material to the website? In the paper is mentioned that result in relighting can be found at relighting_results.html. and i want to look into how good it can be relight. Thank you.

@georgetang50 4 ай бұрын

Please release the code

@bilalse6862 4 ай бұрын

Very impressive paper, thank you guys !

@Copa20777 4 ай бұрын

This is great..❤

@wolpumba4099 4 ай бұрын

*ELI5 Abstract* *Imagine you have a magic drawing machine!* This machine can understand words and make pictures with just a description. But sometimes, the pictures only show the object from one side, like a flat drawing. *We made the machine even better!* Now it can learn from real photos of objects. We taught it how to make a 3D picture inside its head, so you can see the object from any side, as if it were real! *Our pictures are super cool!* We can tell the machine "a red bouncy ball" or "a fluffy brown dog," and it makes a picture that looks just like the real thing. You can even spin the picture around to see it from all angles. It's like magic! *Abstract* Recent advances in text-guided 2D image generation have spurred interest in 3D asset generation. However, existing text-to-3D methods often produce non-photorealistic objects lacking realistic backgrounds. In this work, we present ViewDiff, a method that leverages the power of pretrained text-to-image diffusion models to generate 3D-consistent images from real-world data. Our key innovation lies in integrating 3D volume rendering and cross-frame attention layers into a U-Net architecture. This enables our model to generate views of an object from any viewpoint in an autoregressive manner. Trained on real-world object datasets, ViewDiff produces instances with diverse, high-quality shapes, textures, and realistic backgrounds. Our results demonstrate superior visual quality and consistency compared to existing methods, as measured by FID and KID scores. disclaimer: i used gemini

@faruknane 4 ай бұрын

Great work!

@dangthanhtuanit 4 ай бұрын

I wonder if there is an actual implementation since code is not available?

@adrianstarfinger5721 4 ай бұрын

Impressive work!

@briancunning423 4 ай бұрын

Amazing. Have you tried feeding the images into photogrammetry software or Gaussian Split software to test the consistency of the 3D?

@lukashollein3985 4 ай бұрын

Yes, this works! You can check the figures 17 to 19 in the paper: lukashoel.github.io/ViewDiff/static/viewdiff_paper.pdf

@manu.vision 5 ай бұрын

😮

@couragefox 5 ай бұрын

Really qant to try it. Please let us know if the code will be released...

@stevenlk 5 ай бұрын

wow that’s impressive

@ritikkothari2787 5 ай бұрын

for calculating parameters of a particular layer, we do consider he size oof kernal (including depth) which i guess the professor missed! for example in vgg : for first layer - (3X3X3 + 1) * 64 for first layer and (3X3X64 + 1) *64 for the second layer.

@adrianstarfinger5721 5 ай бұрын

45:17 Here you talk about Neural Rendering, e.g. NERF which is using an MLP, and you call it an implicit function. However, in the first lecture we learned that MLP's are actually not implicit functions.

@ottorocket03 5 ай бұрын

Its funny how you spend hours on watching these videos and learn the material presented in the lecture but the exam is about 3D point clouds and backprogation on convolutional layers

@HootyTooty 5 ай бұрын

I believe the result vector is in the wrong order After solving the corrected linear equation system of (30, 10) = (77) (10, 4 ) = (28) i got (m) = 1,4 (t ) = 3,5