วีดีโอ

Could you please explain why at 7:48 that last term is a constant? Thank you!

Very informative! Step by step approach for OD in CV. Best video so far 👍❤

This explanation is absolutely fantastic.

Amazing video. Probably the best explanation I've seen on the internet. However, I'm still struggling to understand the encoding and statistical explanation of the model. You say that we need to compute p(z|x), but we can't because its computationally intractable, so instead we estimate it using q(z|x). However, my question is firstly: how do we calculate the KL-divergence between q(z|x) and p(z|x) if we don't actually know p(z|x) (at 7:34)? If we did, why couldn't we just use that instead? Next, you say that we sample from the distribution p(z) to generate new pieces of data. This does not make sense to me. Isn't p(z) a standard gaussian? If we sampled from p(z) wouldn't we just get non-sensical, random results? Why don't we sample from the learned distribution of q(z|x) instead? Here's my thought process, please correct me where my understanding is wrong: 1) Imagine we have images of circles that a VAE must reconstruct. 2) We encode it into a 2-dimensional latent space. 3) The decoder decodes the a point sampled from the latent space and generates a new image of a circle. In step 2 we encode images by estimating p(z|x) through q(z|x). Let's say the encoder learns that the two dimensions of the latent space are radius and position. Then, for every single image the encoder finds the latent variables z and turns it into a distribution of radius-position combinations which approximate a standard gaussian (p(z)) (I imagine this looks something like a joint distribution between two normally-spread random variables). We do this enough times and eventually we have a latent space represented by the probability density function q(z|x) which organizes the space of all seen circles into varying spaces within q(z|x) (I imagine this looks like approximately standard gaussian clusters spread out throughout a 2D latent space with axis radius and position where each cluster represents a certain type of circle x). In step 3 we decode these images by sampling from q(z|x) (which is differentiable through the reparameterization trick). Then, we can conduct the reconstruction loss between the generated output x' (or f(x)) and original input x (this makes sense to me) and calculate the KL-divergence between our estimation of the latent space q(z|x) and a standard gaussian p(z) (this doesn't make sense to me). - In this part, why do we take the KL-divergence between these two terms? To my understanding q(z|x) is our prediction of the true latent space p(z|x). If we tried to make q(z|x) as similar to p(z) as possible, would we not just see q(z|x) turn into a giant standard gaussian? Why would we want that? Am I understanding things correctly? (probably not) But more importantly, could you please correct me on what I'm misunderstanding and answer my questions? Again, excellent video. It just seems like there are some kinks which I have yet to work out because of my inexperience.

Man, I love you. You're the person I was looking for. Thanks for great explanation without omitting details and things that can be unclear.

Amazing !!! v8 is coming...

what a clean explanation!!!!!!!!!!

Awesome explanation, Thanks!

please make video Mask R-CNN sir 🙏🙏

can i get the parameter file of this DIT model (trained on mnist) directly?

you have made my literature survey 10 times easier. May I reccomend you to look into transformer and attention based object detection model, starting with DETR. Love your content <3

God thanks

:) thanks!!!

Does YOLOv7 have a similar architecture?

yoyo

Can you please share prediction code

Best explnation of Denoising Diffusion Probabilistic Models!

So much great info in 8 minutes. Thank you so much!

The quality of the writings too poor to see the equations

thank you so much, this video is very helpful to me. you are very generous.

Hey I am a new subscriber can you explain the implementations of LayoutLMV3 and UDOP and help implementing from scratch

Can you explain to me why there is a break on line 114 in the train_torchvision_frcnn.py. It now looks like because of the break it will only use one batch and than break out of the epoch? I really like your videos thankss

I am trying to use this model to train coco, but I am having issues using it, seems the model is veru structured to be trained on PASCAL VOC, any idea how can I adapt it to COCO? great video

If you look at the latent space images at 37:26 you cannot believe the decoder can re-generate the original image from it. As there is simply a lot of missing information. Any explanation on how it does it? First I thought this is due to original image information leaking through through skip-connections between down and up blocks. But we are not using those in the auto-encoder.

In the sampling algorithm (algo 2), I don't understand why we have to add noise z back in. Can anyone explain this to me?

Hi, great explanation of RCNN with very useful insights which often are skipped. I am especially grateful for answering questions like "Why SVM, Why different IOU Thr, etc."

Brother so good keep it up

Can we train our dataset with vq-gan (without transformer) and then use it in train_ddpm_vqve?

Can you explain us how Multi-view Diffusion Base Model works please

Can you please share the notes of all object detection videos

Just what i was needing. Thanks 🙏🏻

Please answer me, how can they train CLASS SPECIFIC bounding box regressor? So they input class as input in one model and regress the bounding box or they build multiple(if they detect 6 class then we build 6 model) model and each model they train on specific bounding box regressor? Please answer me

Please create yolo panoptic sir, it would be a huge help and it has so many applications

It was great! Good luck!

why gussian noise only added. Not Rician, Laplacian etc.. there are so many other probability distribution.

Bro no one in this platform explained clearly as much as you Thankyou for providing these lectures for free of cost . I think for paid courses also no one can explain this much thankyou again

Subscribed

one of the finest videos on yolo available on internet. contains intuitive as well as detailed explanation (right from research paper). concepts like these are hard to explain in so much detail. thanks a lot for the amazing work, cheers!

Fantastic video! You’re undoubtedly on your way to becoming one of the top lecturers in Generative AI. I’m excited to see more of your work in the future!

Thanj you very much for the video, it is very interesting. Though, I have one question, at 15:40 timestanp. You mention that there may be a situation, where ground truth box doesn't have big IoU with any of anchor boxes. How do we pick these anchor boxes (i just cant get which methodology we have to follow when picking the dimensions for anchor boxes)?

How does self attention work in convnets (instead of transformers)? 😊

Great video, but why do you add 1E-6 when calculating your IOU?

what do you mean by topk proposals 2000, is this from the single image we take 2000 proposals?

Thanks for this wonderfully intuitive video! It provided a fantastic breakdown of the fundamentals of diffusion models. Let me try to answer your question about why the reverse process in diffusion models is also a (reverse) diffusion with Gaussian transitions. Why Reverse Diffusion Uses Gaussian Transitions 1. Forward Diffusion Introduces Noise Gradually Remember the β term? In the forward process, β is chosen to be very small (close to 0). This ensures that Gaussian noise is added gradually to the data over many steps. Each step introduces only a tiny amount of noise, meaning the transition from the original image to pure noise happens slowly and smoothly. This gradual noise addition is crucial because it preserves the structure of the data for longer, making it easier for the reverse process to reconstruct high-quality images. If we added large amounts of noise in one go, like in VAEs, the original structure would be harder to recover, leading to blurrier reconstructions. 2. Reverse Diffusion Needs "Gaussian-Like" Inputs The forward process only involves adding isotropic Gaussian noise at each step. This means the model learns to work with samples that are progressively noised in a Gaussian way. However, in the reverse process, when the model predicts the noise at each step, the resulting sample isn't guaranteed to remain Gaussian-like. To fix this, after subtracting the model's predicted noise, we add a small Gaussian noise with a carefully chosen variance. This step helps "Gaussianize" the sample, ensuring it aligns with what the model expects at the next time step. This small added noise smoothens any irregularities and makes the reverse process more stable, resulting in higher-quality outputs. Step-by-Step Noise Removal The reverse process works by removing noise step-by-step, moving from pure noise back to a clean image (closer to x0 ). This gradual approach is crucial because predicting small changes (i.e., removing a little noise at a time) is much easier for the model than trying to reconstruct the clean image in one big jump. This is why diffusion models produce sharper and more realistic images compared to VAEs, where predictions often result in blurry outputs due to the lack of such gradual refinement.

Excellent. Complete, clear and to the point

I am new to this field can anyone provide me with the prerequisites to understand this video

thank you very much!!! great explanation ❤

Thank you! It was amazing. While there are limited content available for diffusion models, you did pretty nice.❤

Excellent Work ! Please Don't Stop :)

Please don't use music in the background, it's very distracting, thanks.