Truly the BEST video for RPN. Huge thanks and congrats to this guy!!!! Other RPN related videos don't compare to this video by far. I heartily request you to make similar videos for the whole Faster RCNN and/or Mask RCNN process; don't worry if it is too long. I would rather spend 2 hours in a single video rather than wasting time and effort in looking for other non-sense videos.
@quarter2018
4 жыл бұрын
Thank you for your recognition and encouragement. Now we are making video about Faster RCNN. It will come soon. I hope that you will like it.
@lovemormus
4 жыл бұрын
hey Bro, just let you now, this is one of the best clips for explaining RPN! Other RPN videos on KZitem do not reveal the details because even the uploader is not familier with rpn, so they just repeat rubish. Nice work!
@quarter2018
4 жыл бұрын
Thank you for your recognition and encouragement.
@lancealexzander7416
3 жыл бұрын
I guess it is pretty randomly asking but does anybody know of a good place to watch newly released tv shows online?
@mosesmoises4009
3 жыл бұрын
@Lance Alexzander ehh lately i have been using Flixportal. just google for it :) -moses
@lancealexzander7416
3 жыл бұрын
@Moses Moises thanks, signed up and it seems to work :) I appreciate it !
@mosesmoises4009
3 жыл бұрын
@Lance Alexzander Glad I could help :)
@Kuro-uc4jp
3 жыл бұрын
This video explains the concept incredibly well! Much better than most generic explanations out there that fails to delve deep into the mechanisms of faster-RCNN, great work!
@quarter2018
3 жыл бұрын
Thanks a lot! We will keep publishing good videos.
@spikewong9740
2 жыл бұрын
The accent of oral English of Taiwanese and Mainland Chinese do have a lot in common which makes your presentation much easier for me to understand, thanks, buddy!
@quarter2018
2 жыл бұрын
I am glad that you enjoy this video ^^
@carlosjaviergarcialopezdeh9213
2 жыл бұрын
As the other comments around, this is the best video explaining RPNs, thanks a lot
@quarter2018
2 жыл бұрын
Thanks for your appreciation.
@yagmuraktas2423
2 жыл бұрын
Best video I have ever seen about RPN!!!! I dont understand this subject for days now all clear thank you a lot!!!! Please keep going with making videos you are the best!!
@quarter2018
2 жыл бұрын
Thanks for your recognition. We will keep going with this.
@sheenanasim
4 жыл бұрын
OMG! this is amazing. So glad I stumbled upon this. Best video on RPN. Thank you!
@quarter2018
4 жыл бұрын
Thanks for your recognition.
@johngrabner
3 жыл бұрын
Best video on this topic. Excellent. Other videos very superficial, your is full of details.
@quarter2018
3 жыл бұрын
Thanks for your recognition. You are so welcome
@truongg2340
3 жыл бұрын
this is the best video in RPN so far!!!! thanks so much for your significant level of detail!!
@quarter2018
3 жыл бұрын
You are so welcome
@farvezfarook3422
2 жыл бұрын
Nice Explanation. Very well articulated. Thank you
@quarter2018
2 жыл бұрын
You are so welcome
@akintundeoladipo95
4 жыл бұрын
Loved this explanation and I will be sharing it with colleagues!
@quarter2018
4 жыл бұрын
Thanks very much.
@vacous1994
3 жыл бұрын
This is the best RPN explanation on KZitem (if not for all resources). Thanks a lot!!!!
@quarter2018
3 жыл бұрын
You are so welcome ^^
@kk-mh8uf
4 жыл бұрын
Great job. I've never come across any simple and clear explanation of RPN like this one before.
@quarter2018
4 жыл бұрын
Thanks a lot. We also provide the video about Faster R-CNN based on RPN for object detection. We wish you will like it.
@haoxiong7077
3 жыл бұрын
watched 10 videos related to the original paper, only this one gave me a full understanding of what is RPN
@quarter2018
3 жыл бұрын
Thanks for your recognition. We have sent the slide to your email. I wish you will like it
@korinlifshits8780
3 жыл бұрын
The only lecture that really helped me understand what is going on!!! Can you upload an explanation on mask-rcnn would be really appreciated. Thank you a lot!
@quarter2018
3 жыл бұрын
Thanks for your recognition. Mask R-CNN is the one in our schedule ^^.
@satishy1
3 жыл бұрын
you did an amazing job..made it so simple..this is the best explanation
@quarter2018
3 жыл бұрын
Thanks a lot ^^
@deeptijoshi6118
4 жыл бұрын
This is amazing..his is one of the best explanations of RPN, I have come across..Well done..amazing job!!!
@quarter2018
4 жыл бұрын
Thanks a lot. We also have a video about Faster R-CNN that using RPN for object detection. We wish you will like it.
@songxiaohu6435
3 жыл бұрын
Really the best explanation of RPN, thanks a lot!
@quarter2018
3 жыл бұрын
I am glad that you like this lecture. Thanks for your recognition.
@karthikeyasai4683
4 жыл бұрын
truly detailed view on rpn, by far the best on rpn.. thanks so much
@quarter2018
4 жыл бұрын
Thanks for your recognition.
@muhammadzubairbaloch3224
3 жыл бұрын
Great lectures.
@quarter2018
3 жыл бұрын
Thanks for your recognition.
@whasuklee
3 жыл бұрын
Superb! Thank you so much for the great explanation! 👍👍
@quarter2018
3 жыл бұрын
Thanks for your recognition.
@surayuthpintawong8332
3 жыл бұрын
Thank you so much. This is the clearest explanation about RPN! Am I understand it right that if we want to get the coordinate of the predicted bounding box, we have to get it from the predicted vector d (by inversion).
@quarter2018
3 жыл бұрын
Your understanding is right.
@Banefane
Жыл бұрын
Thank you very much :)!
@quarter2018
Жыл бұрын
You are so welcome ^^
@yasminemasmoudi9963
Жыл бұрын
Thank you !
@quarter2018
Жыл бұрын
You are so welcome ^^
@marciogaraz
3 жыл бұрын
Perfect!!
@quarter2018
3 жыл бұрын
Thanks a lot! You are welcome
@danbrown6698
2 жыл бұрын
Although your oral English is pretty poor, your content is fairly clear
@quarter2018
2 жыл бұрын
First, thanks for your comment "your content is fairly clear". For your suggestion, what we currently do is to provide subtitle. In the future, we will spend more time on how to improve our oral English. Thanks for your comments.
@prajvalp2868
4 жыл бұрын
Amazing explanation. Thank you for the video!! Is there any basic code available to implement the network ?
@quarter2018
3 жыл бұрын
Sorry, we still work on this.
@ryancusack5909
2 жыл бұрын
Im not sure I understand the anchor labeling and anchor sampling step in training. When you first select a random image do you generate random anchors throughout the image at random locations? Why don't we use the ground truth bounding box in training instead of the highest IOU anchor box with the ground truth.
@quarter2018
2 жыл бұрын
(1) The anchors are fixed at each point of the conv. feature map but not randomly generated. (2) During training, you have to compute the difference (called loss) between ground-truth bounding boxes and predicted ones. The selected anchors are considered as the predicted bounding boxes for computing the loss.
Great explanation, well done, the best I have seen so far. I have a question. For a good RPN training how many epochs we need? 60000 with 256 batch size? Thank you
@quarter2018
4 жыл бұрын
Thanks a lot. How many epochs you need is significantly depend on the network parameters and your dataset.
@romanasalim5335
Жыл бұрын
sir..could you plse share the ppt link to download.U have provided the best explanation of RCN.
@quarter2018
Жыл бұрын
You can download our slide in pdf format from our web-site via the link gg.gg/quarter
@jaydenlu764
4 жыл бұрын
很棒
@quarter2018
4 жыл бұрын
謝謝你的肯定
@vishakraj5377
3 жыл бұрын
Sir, thank you for your video, I learned easily, I have one question - in RPN architecture(6.07), we attach 9 anchors centred at each point in conv. feature map and predicting 6 parameters (dx, dy, dw, dh) and (Pobj, Pbg). This is actually means that we are applying 9 anchors on every (3×3) sliding window in conv. feature map and predicting the 6 output parameters? If it is, then does it looks like we are applying 9 different sized sliding window on the conv. feature map instead of 3×3 sliding window... Could you please clarify my doubt, thanks
@quarter2018
3 жыл бұрын
We attach 9 anchors at each point in conv. feature map. Then, we predict 6 parameters (dx, dy, dw, dh) and (Pobj, Pbg) for each anchor. By using anchor with its predicted 6 parameters, we can produce an object proposal (using regression parameters (dx, dy, dw, dh) ) with its probability belonging to object or not (using (Pobj, Pbg)). In other words, we will produce 9 object proposals for each point in conv. feature map. You can imagine that there 9 different-sized sliding windows (named as anchors) used for producing 9 object proposals at each point in conv. feature map.
@vishakraj5377
3 жыл бұрын
@@quarter2018 Sir, thank you, I understand:-)
@ivanphillips8529
3 жыл бұрын
Great explanation, thank you! I have one doubt: how do you discard anchors which are not included into 256 selected anchors for mini-batch? I think we cannot just ignore them. In training process, depth of output class array is fixed (in your case it is 18). For example, for one point of features map, if one of 9 anchors is considered as foreground [1, 0], and one is considered as backgorund [0,1], how to label other 7 anchors? How to prevent RPN to backpropagate loss over that 7 anchors?
@quarter2018
3 жыл бұрын
Thanks for your recognition. The answers to your questions are as follows: (1) In general implementation, we randomly select 128 positive and 128 negative anchors for a mini-batch. (2) For a mini-batch, only the selected 256 anchors but not all the anchors (9 anchors x num. of feature points) will be used to update the parameters of RPN by back-propagation. The loss of each selected anchor is computed independently and the final loss is obtained by summing all anchor losses. The detail of how to compute the loss of an anchor is at 16'26'' of the video.
@ivanphillips8529
3 жыл бұрын
@@quarter2018 Thank you very much for your answer. If I understood correctly, shape of input data for training process will be [256, 3, 3, depth of features map] and output for classes will be [256, 1, 1, 18] ? For example, if only 1st of 9 anchors is foreground, output for that point will be [1,0 , 0,0 , 0,0 ... 0,0] with length of 18? Please correct me if I am wrong. Thank you.
@quarter2018
3 жыл бұрын
In my understanding, the input of RPN is [W, H, feature depth] and output will be [W, H, 9*4] (regression map for 9 anchors)and [W, H, 9*2] (class map for 9 anchors). If only 1st of 9 anchors at the point (x,y) of the feature map is foreground, output for that point (x,y) will be [1,0 , 0,0 , 0,0 ... 0,0] with length of 18.
@ivanphillips8529
3 жыл бұрын
@@quarter2018 I have problem with regression. RPN backpropagate error (loss) over all of cells (W*H*9*4). I don't understand how to focus RPN on backpropagating loss over only positive anchors (positive anchors have [x,y,dw,dh] values, while other anchors are set to [0,0,0,0]).
@quarter2018
3 жыл бұрын
After anchor labeling, we assign a label to each anchor. So you only have to take the anchors labeled as positive for computing the regression loss in implementation.
Пікірлер: 85