You explained it really well. Big thank you. But in the recent modification of the the paper, the author changed the FCN to FPN (Feature Pyramid Network).
Thanks. Will do. Working on another video on various Convolution Neural Net Architectures. I'll have that up in a few days. It's going to be a new kind of video, but I'd consider it "stuff like this". So stick around :)
At 3:48, how exactly does max pool rotational invariance?? I understand translational invariance but a rotation would make different features activated
If i want to use pretrained R-CNN for my own dataset to segment ( delineate) background from foerground , do i need to annotated or label my data ? The data i am using if person image ..
Your video is very good! Ask me a question, what would be the variables or conditions that I should consider when defining the variable STEPS_PER_EPOCH? Because I have a dataset with 50 images.
Steps per epochs is the data size divided by batches, but in a rounded sense: if your batch size was 25, you would have 2 steps, but if your batch size was 24, you would have 3 steps, one for the two images that are leftovers after the batches have been created. And the thing is, there is no "rule of thumb" when deciding the batch size - it is more theoretical, because bigger batches imply that your weights and biases will be updated less often in one epoch so it is easier for your computer to do, but smaller batch sizes contribute to the precision of the model since they act like a regularization. I would go with 25 steps, so batch of two, in your case. I use 64 or 128 when working with millions of inputs. But the great thing is that your small dataset can be made better by using image augmentation - it is a built in tensorflow function for that, it will flip your images at random, rotate them, crop them, making your dataset seem larger than it is because, if you just use the flipping option, your one image can be seen as 4 different images in the input. It is important that, if you are doing segmentation, you apply the same augmentation on your "gold data", or the manually created masks and segmentations that are used as true output, one you compare your predictions to.
I said "analogous to the One-Vs-Rest approach". It is a method of multiclass classification where we construct K (number of classes) binary classifiers. Each classifier determines whether a sample belongs to class k or not i.e. "one" Vs "the rest". I use it in this context to represent the construction of 3 binary masks (human, dog, cat). Thanks for watching Ha Nguyen! Stick around for more content :)
At first thank you very much for this video. Your videos quality are very good. I have started to watch your videos. Can you Using Mask RCNN we can detect human class, from that human class can we detect human face ? Then which algorithm will i use to detect face ? Can you please give me some suggestions. And is it possible to use same dataset for human detection along with face detection ??
Very well... thanks for the video.. I had some difficulty completely understanding how ROIalign eliminated mis-alignment.. I understand better now... Thanks
Thank you for taking the time and efforts to make this video. Side note: the creepy whispered "subscribe" at the end of the video has more of a repulsive effect and doesn't really make me want to subscribe (more like making me want to close the video as fast as possible). The positive energy given during the video would probably work a lot better if it were used to ask for subscription too.
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
![Hopfield network: How are memories stored in neural networks? [Nobel Prize in Physics 2024] #SoME2](/img/n.gif)
you explained what all the others didn't. Thanks a lot now all the dots are connected in my mind.
You explained in a very simpler way. A big thank you from my side. All the best for your upcoming codeEmporium.
You explained it really well. Big thank you. But in the recent modification of the the paper, the author changed the FCN to FPN (Feature Pyramid Network).
bro you're doing such a great job. your videos are so helpful.
what a great video!!!
great exploration
just started learn a computer vision, for me this video is the most understandable
Thanks Man. You are a beast in explaining, everything is perfect.
This is great! Please keep on making stuff like this xD.
Thanks. Will do. Working on another video on various Convolution Neural Net Architectures. I'll have that up in a few days. It's going to be a new kind of video, but I'd consider it "stuff like this". So stick around :)
At 3:48, how exactly does max pool rotational invariance?? I understand translational invariance but a rotation would make different features activated
Nice explanation especially on the ROI align part! I understood based on your explanation!!! Thanks!
Thanks for your explanation! It saves me from the complicated explanations of my lecture.
This channel is so legit good omg
Thank you! Simple and clean thought process :)
Thank you so much. Very nice Introduction and Explanation. I understood a lot even though I lack a proper background in computer vision!!
Awesome!! Thanks for watching!
Subbed this is a really really well made easy to understand video. Hope to see more from you in the future!
very much lucid explanation. I would request you to make a detailed video on the subtopic discussed here ROI,ROI pooling and ROI align
Thanks a ton the the compliments. Maybe a future video? I need to motivate it more generally if I’m going to make a video on it. So possibly:)
@@CodeEmporium
Yes, requesting you to nailed it 😅.
Great explanation 👍🏻👍🏻
Thank you for this, really good explanation and straight to the point
wow boy, this is a REALLY GOOD video. Thanks!
Can this masked rcnn be used for overlapping leaves with diseases???
Great content and able to understand the concept in a very little time
your video is very helpful and to the point.thank you very much
Good summary and ROI ALIGN description.
Great explanation, thanks a lot! Can I ask what you mean when you say "when computing the mask, a loss of KM squared is incurred" at 6:44?
the time complexity to compute all the masks for M*M region of interest for k possible classes is k*(M)²
Thank you great work! Is there an easy (beginner friendly) explanation how ROI align works?
Great explanation!
Great video! Thank you
hello can you also explain ho to plot graphs on mask rcnn demos
Gr8 work dude.Subscribed
Thanks !
Very detailed video. Thank you very much.
Welcome!
very nice explanation. Thanks
Thank you for explanation how do i save the model ?
Great Explanation, will follow your videos! Thanks for the share
Thanks Tiago Freitas. Glad to know you are on board!
Nice work man!!!!
Thanks! As long as it's useful!
Excellent video!
James Thanks! So glad you liked it !
Great video, keep rocking.
Nice, You made it look easy!
That's what I was going for. Research papers make everything complicated. Why not change that ;)
If i want to use pretrained R-CNN for my own dataset to segment ( delineate) background from foerground , do i need to annotated or label my data ? The data i am using if person image ..
What do you mean by pixel to pixel alignment?
Your video is very good! Ask me a question, what would be the variables or conditions that I should consider when defining the variable STEPS_PER_EPOCH? Because I have a dataset with 50 images.
Steps per epochs is the data size divided by batches, but in a rounded sense: if your batch size was 25, you would have 2 steps, but if your batch size was 24, you would have 3 steps, one for the two images that are leftovers after the batches have been created. And the thing is, there is no "rule of thumb" when deciding the batch size - it is more theoretical, because bigger batches imply that your weights and biases will be updated less often in one epoch so it is easier for your computer to do, but smaller batch sizes contribute to the precision of the model since they act like a regularization. I would go with 25 steps, so batch of two, in your case. I use 64 or 128 when working with millions of inputs. But the great thing is that your small dataset can be made better by using image augmentation - it is a built in tensorflow function for that, it will flip your images at random, rotate them, crop them, making your dataset seem larger than it is because, if you just use the flipping option, your one image can be seen as 4 different images in the input. It is important that, if you are doing segmentation, you apply the same augmentation on your "gold data", or the manually created masks and segmentations that are used as true output, one you compare your predictions to.
Very well explained....can you please elaborate the mask branch with pixel values
Isn't Object Detection + Semantic Segmentation = Panoptic Segmentation?
nice explanation. subbed
Thanks Mark! Been following your channel as well. Interesting stuff.
thanks! glad to see more channels making videos on the subject.
@@MarkJay Quality content creators!!!! Thank you guys!!!
Fantastic
Does it apply to orbit semantic segmentation?
how to prepare own dataset for this I dont want to use cocodataset
thank you
Please make a video related to visual question answering
At 6:41 what is "analog is 2 a 1 versus rest approach"? Thank you very much.
I said "analogous to the One-Vs-Rest approach". It is a method of multiclass classification where we construct K (number of classes) binary classifiers. Each classifier determines whether a sample belongs to class k or not i.e. "one" Vs "the rest". I use it in this context to represent the construction of 3 binary masks (human, dog, cat). Thanks for watching Ha Nguyen! Stick around for more content :)
Tự động đánh dấu phân biệt sắp xếp vào những người và điểm thường lui tới vào kho
explain very easy! thanks
Anytime
that’s impressive 😍
It is indeed :)
you should explain ROI align in more mathematical detail
great vid
Thanks!!!
I want to classify body movements. What are your ideas?
I'm also in a similar research, if you could found related details please let me know, my email is samitha156@gmail.com
Great video
Nhà thông minh của trí tuệ nhân tạo🙂
Awesome. Thanks!!
Just found another great tutorial on AI
Why - thanks for the kind words ;)
At first thank you very much for this video. Your videos quality are very good. I have started to watch your videos. Can you
Using Mask RCNN we can detect human class, from that human class can we detect human face ? Then which algorithm will i use to detect face ? Can you please give me some suggestions. And is it possible to use same dataset for human detection along with face detection ??
preciate you stay blessed
Do u know where I can find a code for it
link in the description
Thanks :)
Đánh dấu địa điểm thường xuyên đến
Thanks! \m/
Can I please get the ppt?
Amazing video
Thanks! These aren't actually slides. I create these slides in my video editor directly.
Very well... thanks for the video.. I had some difficulty completely understanding how ROIalign eliminated mis-alignment.. I understand better now... Thanks
Glad it helped! Really sorry I can't help you out with the slides though.
its ok.. thanks
Thank you for taking the time and efforts to make this video.
Side note: the creepy whispered "subscribe" at the end of the video has more of a repulsive effect and doesn't really make me want to subscribe (more like making me want to close the video as fast as possible). The positive energy given during the video would probably work a lot better if it were used to ask for subscription too.
Thank you for the explanation!! Can you share me your slides?
Kyaaa bat hai
非常好
Thu thập thói quen hành vi người dùng hay đi qua chung một tuyến đường của trí tuệ nhân tạo
You give very vague overview, no insights into how the training is done and all.
I love your non indian accent
Don't put ur scary face