Feature Pyramid Network | Neck | Essentials of Object Detection

This tutorial explains the purpose of the neck component in the object detection neural networks. In this video, I explain the architecture that was specified in Feature Pyramid Network paper.

Link to the paper [Feature Pyramid Network for object detection]

The code snippets and full module implementation can be found in this colab notebook:

The torchvision has a more flexible implementation which would take more than 3 feature layers from backbone

[paedrufernando]

Keep the pearls of wisdom dropping sir..Privilage to learn from you miles across...

[lostpenguin]

very helpful! I really like that you're explaining it with an example with concrete numbers!

[dopnhbi]

Sir, I have a lot of to say after finding your video on YouTube but just ❤, respect and thank you. 🙏🙏

[brunodias]

I am so happy I found this video. Really good content!

[NehadHirmiz]

Excellent tutorial. Thank you very much.

[science.]

is useful to add channel and spatial attention in conv layers to improve

[applestarpie]

I like your videos, which are easy and fun to learn. Thanks a lot!

[rampavanmedipelli]

Thank you... excellent clarity... please try to make a tutorial on anchor free detectors like FCOS..

[dmgeo]

How is this different from U-net? I think they're pretty similar if you think that in the U-net you're going down in the encoder, up in the decoder and sideways with the skip connections. It's like an upside-down U-net

[ranjithtevnan]

I have 2 questions. How are the 1X1 and 3X3 CNN used trained to obtain the weight parameters? Also shouldn't 3X3 with stride 1 change the dimension, though it keeps the number of channels the same the size of the output feature would have changed and reduced by 2

[vincentpelletier]

I don't know if I got this wrong but if I take a 1x64x26x26 feature through a convolution that has a K=3 and S=1, I will definitely not end up with a 1x64x26x26, but with a 1x64x24x24. To achieve the desired shape would require a P=1.

If I'm not correct, would someone please explain how the dimensions would work in this case?

[krishnachaitanya]

This is quite informative and helpful. Can you please create a video on prediction heads in fpn as in how to assign a predicted bbox to a particular feature map. That would be quite helpful.

[yogeshwarshendye]

If done with UNet, it won't require upsampling as we concatenate the layers right?

[kylehuang]

Could you give a tutorial of diffusing model to your VAE series? Its related and would like to see your explanation!

[LongLeNgoc-qqqn]

what about height and width are odd number (415), sir? In that case, the size after conv and after upsample is miss match. How to fix that, please!

[DIAHAYUNINGTYASWATI]

Do you know how to combine AFPN with the YOLO v8 algorithm? If you know, please tell me. Thanks

[rampavan]

Could you give a tutorial on the vision transformer model for object detection?

[ufmdubj]

Thanks a lot! would be the following videos soon?

[lordfarquad-bydq]

thank you for the content, next video soon?

[cheeziobodini]

Instead of doing the upsampling via pytorch module and being angry about it, would it be any more useful to train an additional layer to do the upsampling instead? I'm thinking of a layer analogous to the decoder layer in an autoencoder.