Please do! I'd love to see it. Also, could you do an implementation video where you go over a new paper and explain your thought process on how to implement it at the same time? What are the alternatives for each of your decisions, how do you go about solving errors, etc? I haven't found any video where I could observe an ML engineer's process of reading and implementing a paper from start to finish (with all the debugging, researching, refactoring, testing, evaluating, etc.). I would watch it, no matter how long it is. And I believe it would be very valuable.
I like your intuition on the volume rendering. It helped me understand this topic. However, the description of T(t) is a bit misleading IMO. You say it is "How much light has been blocked up until point t." It's actually the opposite of that. My description would be "How much light remains up until our current point". If it's a large number, we have a large amount of light left and we will use a lot of the color when rendering. If it's zero, ALL the light has been blocked. Hope this helps some others.
Alright. This is very interesting for sure, like basically you do 3D reconstruction with Neural Networks... However, I don't really see use cases of this based on the fact that we need to do this for each particular scene. This can be easily implemented with the 3D reconstruction algorithms in practice.
Yeah that’s the thing. Yes you do need special hardware but I’m happy with the results from my iPhone lidar scanner. No training needed and fast. Same or better results
Thank you for the great explanation. It would be awesome if you could make the NeRF from scratch video to better understand the formulas. Even a basic implementation would be great.
At 16:30-16:45 where you struggled to explain why neural networks work better with high dimensional data. Am I correct to say that neural networks need high dimensional inputs in order to avoid linearity, which will make the gradients vanish resulting in the networks stop learning? Btw great video explaining the concept in as simple manner as it could get for convoluted topic such as NeRF. Subscribed! :)
Hey buddy, is there any field in DL that you have not put your head in ?! lol. Whatever new thing I begin and look for some tutorials, your videos keep popping up. Wish you all the best.
Still I didn't understand how the training works. So, is it like we build dataset of lines (rays) and then train MLP on that? Can you please explain and show how to do it?
Hey Aladdin! love your videos. I made 2 projects related to deep learning and want to start replicating papers. I have some proficiency in Tensorflow (2.0). Wanted to ask how do you come about replicating papers i.e. how do you find these papers and start replicating them. Also, If I have replicated a paper, should I add this as my "portfolio project".
Does anyone know, how does NeRF know what is occupancy of any point along ray? does it just minimizes the loss and thus learns both corresponding sigma and Rgb of the scenes ? To be honest, can not understand how does MLP learn the distance and corresponding RGBsigma of each point on the ray, intuitively it should know some sort of 'depth' of the scene(maybe through the other view pictures)?
A little thing, I'd like to request......... Please don't use the white background. It hurts the eye, battery, and pretty much all programmers use dark theme anyways.
Thinking about doing a NERF implementation from scratch in PyTorch next
Awesome! Would be very keen on a tutorial
Please do! I'd love to see it.
Also, could you do an implementation video where you go over a new paper and explain your thought process on how to implement it at the same time? What are the alternatives for each of your decisions, how do you go about solving errors, etc? I haven't found any video where I could observe an ML engineer's process of reading and implementing a paper from start to finish (with all the debugging, researching, refactoring, testing, evaluating, etc.). I would watch it, no matter how long it is. And I believe it would be very valuable.
Looking forward to it, awesome content as always! Does the paper tell how much photos of a scene nerfs need to get a good 3D reconstruction?
Damn! That would be awesome!
Definitely, can't wait! Great intro to NERF, thank you for your explanations :)
This is like the best resource i've found on this topic, would love the implementation from scratch
That means a lot, thanks
please i am having issue running it. please can you help out
I like your intuition on the volume rendering. It helped me understand this topic. However, the description of T(t) is a bit misleading IMO. You say it is "How much light has been blocked up until point t." It's actually the opposite of that. My description would be "How much light remains up until our current point". If it's a large number, we have a large amount of light left and we will use a lot of the color when rendering. If it's zero, ALL the light has been blocked. Hope this helps some others.
I am so glad that I've found your's channel. Started from zero, but now I am working on the StyleGAN on my own :)
Alright. This is very interesting for sure, like basically you do 3D reconstruction with Neural Networks... However, I don't really see use cases of this based on the fact that we need to do this for each particular scene. This can be easily implemented with the 3D reconstruction algorithms in practice.
Yeah that’s the thing. Yes you do need special hardware but I’m happy with the results from my iPhone lidar scanner. No training needed and fast. Same or better results
Perhaps there are missing data or need to create a scene from synthetic data.
If you have all the data, sure other photometry methods work better.
Thanks, Aladdin! It would be interesting to see your implementation :)
A code implementation would be awesome!
Thank you for the great explanation. It would be awesome if you could make the NeRF from scratch video to better understand the formulas. Even a basic implementation would be great.
At 16:30-16:45 where you struggled to explain why neural networks work better with high dimensional data.
Am I correct to say that neural networks need high dimensional inputs in order to avoid linearity, which will make the gradients vanish resulting in the networks stop learning?
Btw great video explaining the concept in as simple manner as it could get for convoluted topic such as NeRF. Subscribed! :)
Would love a code implementation from scratch!
I am new to deep learning should i listen to this videos or go to deep learnng playlist and what is the next step i should folow?
An implementation video would be fantastic!
It would be very interesting to see this from stracth implemented man! :)
Thanks for your videos, your explanations are simple to understand and best!
when you are implement NeRF implementation in pytorch. Because this is time of research in NeRF
did you end up doing the pytorch implementation? If not, please do, it would be great, thanks!
absolutely amazing explanation. Thank you.
can you please make a video for the implementation
Thanks for the simple explanation👏
Nice and simple explanation. Thanks
Hey Aladdin, if you did a paper walkthrough and implementation of GFP-GAN, you would be a god amongst men.
Kindly implement the NERF . thanks
very nice explanation on Nerf in such an easy way
Hey buddy, is there any field in DL that you have not put your head in ?! lol. Whatever new thing I begin and look for some tutorials, your videos keep popping up. Wish you all the best.
Trying to learn this, understanding and experimenting with the Fourier features is much easier with 2D images.
Thank you for making such a great video :)
Great video!!! Could you give us a NERF implementation from scratch in PyTorch?? Please!!! It would be greatly helpful to a graduate student!!!
please it helps us a lot.. can you implement it from scratch
This is amazing. Please do slam with nerf paper explanation
Thanks
Let's say we put a camera at each angle of the room.
We will soon be able to create a 3D reconstruction of a video... That's something !
Nice video. Very helpful. Thanks
Can you talk about BEV and Occupancy Network?
here x,y,z is the camera location right ? not pixel location
Still I didn't understand how the training works. So, is it like we build dataset of lines (rays) and then train MLP on that? Can you please explain and show how to do it?
Yeah, that would be great if you implement it in pytorch
Is (x,y,z) a coordinate of a point in the scene or coordinate of a viewer?
Hey Aladdin! love your videos. I made 2 projects related to deep learning and want to start replicating papers. I have some proficiency in Tensorflow (2.0). Wanted to ask how do you come about replicating papers i.e. how do you find these papers and start replicating them. Also, If I have replicated a paper, should I add this as my "portfolio project".
Hey Aladdin, could you please implement this from scratch?
pls work on the scratch implementation
Nice!
Would love to implement of the paper too
Hey Aladdin, Could you make a video tutorial about TE-GAN ? This is for thermal image enhancement. Good luck and thx.
Does anyone know, how does NeRF know what is occupancy of any point along ray? does it just minimizes the loss and thus learns both corresponding sigma and Rgb of the scenes ?
To be honest, can not understand how does MLP learn the distance and corresponding RGBsigma of each point on the ray, intuitively it should know some sort of 'depth' of the scene(maybe through the other view pictures)?
Hey Aladdin, think implementing RetinaNet from scratch !
Good Job bro
how accurate are nerfs
please explain about the creation of the dataset. How poses will be created from my own scenes?
You can use COLMAP.
A little thing, I'd like to request......... Please don't use the white background. It hurts the eye, battery, and pretty much all programmers use dark theme anyways.
💕 𝔭𝔯𝔬𝔪𝔬𝔰𝔪
Please share your GitHub profile to follow you. 😊, Interesting Presentation !