2014 Fully Convolutional Network (FCN) Paper summary

[@danielkusuma6473]

Thanks for the video! I had difficulties while reading the papers but you break it down really nicely!

[@armagaan007]

Yeah, I was really disappointed. The authors should have used more figures.

[@ILovePianoForever]

very informative and better to understand than the original paper. Thank you!

[@SofieSimp]

Please continue doing this type of video.

[@yashanand6311]

thanks for this video. you explained so well to understand this paper easily!!!

[@phongtrangia9031]

thanks for your summary, this thing is very strange for me but you help me understand and image it

[@tahacharehjoo9473]

Thank you very much! It helped me a lot to understand the paper!

[@vladyslavkutsuruk6432]

You are doing amazing job with these papers explanations, thank you =) Btw, would be great to see your summary on Mask R-CNN.

[@hnull99]

Very nice video and explanations thank you

[@souadyahiaberrouiguet1285]

Thank you for your explanation

[@Maciek17PL]

Your explenation of convolutionlizatiobn is totally wrong because convolution is scaling down the image by constant factor so when using 1x1 convolution final vector will also be of arbitrary size not always 1x1xD

[@usmaniyaz1059]

Hi! Can you make a video on the 'Meta Pseudo Labels' paper. Your paper summary videos are awesome

[@akshayv2849]

Hello sir that was an amazing explanation. I'm currently doing mechanical engineering bacherlor degree and would love to work on Autonomous vehicles and work on the software side of things. Do you have any career guidance.

[@sarynasser993]

thank you great explanation

[@CrypticPulsar]

Thank you for this video!

[@tm-jw2sq]

講得太好了，謝謝你!!

[@samarthshah8498]

Hey, great video and amazingly explained. One doubt: At 15:10, you say that we are now not restricted to flatten it and if we put a bigger resolution image. It will automatically fir to these value at the cost of some loss in information. My question is how do you restrict it?, like pooling cannot restrict it. What is the layer that actually makes it stick to that dimension. Also some place i saw that its not always 1x1xclasses, i saw 7x7xclasses. So what is the logic behind that? I am still new and highly interested in topic so these doubts might not be good ones, but if you can please clear it

[@gianluca3131]

So, take what I'm saying with a grain of salt because I'm not a Professor, but from what I understand of the paper you don't actually get layers of the same size if you use larger images. Let's say you pull the net with 64x64x3 (rgb) images. If you run larger images in testing, where you would have gotten a 1x1x21 layer (we want to classify 21 classes in this problem) you would now have an MxMx21 layer, where M is a size that depends on the combination of layers you used. The point, as I understand it, is that if you do NOT use fully connected layers, replacing them with Convolutional Layers, having layers of different sizes (M in my case) is not a problem, because you just apply the filter you have to all the pixels. What must remain the same is the detph, so in our example 21, but that is determined by how many filters we use in each pass; if we use 21 filters in the last pass, whether the image is 64x64 or 500x500, the depth will always be 21. So if we use a Fully Convolutional Network we don't need an operation like Flattening, because we don't care about forcing all the values in a certain size (in a 1d array for example). Regarding the sense of having 1x1xclasses or 7x7xclasses as output, the sense I think is that if you have an image output, so instead of 1x1xclasses for example MxMxclasses, you can generate a heatmap that tells you in which zone of the image a certain class is present. If you output a 1d array (1x1xclasses), the network can only tell you what class the image belongs to, if it's a volume (e.g. 7x7xclasses) it can tell you what zone has what class. I hope I didn't say any bullshit, if someone more experienced would like to confirm or deny what I said I would appreciate it.

[@kdubovetskyi]

@@gianluca3131 during whole the video I thought that the model uses some global pooling to get the bottleneck, because the author said that we get 1x1xD-shaped bottleneck for *every* input size. Your interpretation sounds more reasonable, I'm glad I found your comment

[@gianluca3131]

@@kdubovetskyi glad you found that useful, I hope I haven't said anything wrong 😄

[@chaimaaderda4655]

Amazing explanation thank you

[@heikeneubau7064]

Thank you very much!!!