Faster R-CNN PyTorch Implementation

[@Explaining-AI]

*Github Implementation* - github.com/explainingai-code/FasterRCNN-PyTorch

[@hieuhienhoa29]

Wow, you explain things in a very detailed and informative way. Thank you! Keep doing this, I look forward to seeing an entire series done.

[@Explaining-AI]

Thank you for this comment and your support.

[@sushilkhadka8069]

Your Channel is a gem. You make concepts so easy to understand. Hats off to your for all the effort. I know you're very busy but when can we expect MaskRCNN implementation? cause that's the goto model for segmentation tasks even while comparing with ViTs. Thanks again.

[@Explaining-AI]

Thank you for these words of appreciation :) I am planning to do couple of videos in generative space(control net is next) and then I will start a segmentation series(including instance segmentation) in parallel with ongoing detection series. And in there I am planning to cover FCN/Deeplab and then will come MaskRCNN . But because these videos take time, I would only be able to work on mask rcnn video about 2 months from now.

[@sushilkhadka8069]

@@Explaining-AI Oh Okay, thanks super excited!! Could you please mention your medium/tds link? Cause I don't want to miss anything you teach.

[@Explaining-AI]

@@sushilkhadka8069 The only place where I post(atleast as of now) is on KZbin. So this channel will always have everything that I create :)

[@mohammadtarek872]

Your video is very helpful for the implementation. I have confusion about the target, why the target is 1x6? Can you please explain?

[@Explaining-AI]

Thank You! The target shape that I mentioned is just the shape of target for a random image present in the dataset. And that image had 6 ground truth objects in it. Thats why @7:27 the target dictionary has labels as Batch Size x Number of objects(1x6), and box coordinates as Batch Size x Number of objects x 4(1x6x4)

[@aasiasalahuddin2795]

can this code be implemented on Jupiter notebook. As you have created different files for training, model building and testing

[@samt5682]

Do you think that faster rcnn architecture is similar to SSD 1. SSD predicts classes and bounding box coordinates (scaled and adjusted w.r.t predefined anchor boxes). 2. The only advantage of SSD is it does in one go and many middle layers together make over 8000 anchor predictions.

[@Explaining-AI]

Faster RCNN(original paper version) actually does not use multi scale feature maps(unlike SSD). It was in FPN paper, where Faster RCNN was wrapped with a FPN and then different scale feature maps were used to learn to predict different sized objects. Faster RCNN + FPN(this is the one used in pytorch implementation as well) is indeed similar to SSD(both using multi scale feature maps to predict objects of different sizes). One difference like you said is that SSD goes from anchor -> predictions through its extra layers whereas FasterRCNN + FPN follows the Faster RCNN way of Anchors->Proposals through RPN and then Proposals->Predictions. Additionally in SSD the first scale feature map that is used is of a later stage(Conv 4_3 of VGG), whereas in FPN, the first feature map used had a stride of 4. From FPN Paper - "to avoid using low-level features SSD foregoes reusing already computed layers and instead builds the pyramid starting from high up in the network (e.g., conv4 3 of VGG nets) and then by adding several new layers. Thus it misses the opportunity to reuse the higher-resolution maps of the feature hierarchy. We show that these are important for detecting small objects"

[@samt5682]

@@Explaining-AI Thanks got it.

[@AlbieT-yw2jj]

Amazing video, thank you for this rigorous design breakdown. You provided lots of intuition that I need to implement Faster RCNN on my own. Especially looking forward to your Part II - Pytorch RCNN since my application uses that module but requires a lot of changes on my part to handle a completely different dataset. When can we expect that video to be released?

[@Explaining-AI]

Thank you for the kind words :) I think I should be able to release it by Tuesday.

[@Explaining-AI]

Part II of this is now published - kzbin.info/www/bejne/j3Kac6Sfirl2ra8

[@mateuszadamowicz9825]

Hi! Can you help me with some sources or code for implementation of MaskRCNN based on this implementation of FasterRCNN. I have troubles with the tensor shapes of data while calculating mask losses and MaskeHead of its own.

[@Explaining-AI]

Hello, this code I implemented very closely follows the official pytorch version so for mask rcnn also, the pytorch's mask rcnn code will have significant overlaps with this version. In the next video of this series, I go through the pytorch's(official) implementation for FasterRCNN, working through all the shapes. Though I dont cover the mask part, but everything thats common between faster & mask rcnn is covered. Maybe just look at it for couple of mins and see if it has what you are looking for. kzbin.info/www/bejne/j3Kac6Sfirl2ra8

[@CodeWithZeyad]

Thank you so much, i have been trying to get into object detection for a while jumping from a research paper to another but i was struggling to find out the acutal implementation but then i found this video which was very well made and made everything finally clear Thank you so much! tho kinda got a curious question :p , how long did it take to make that video?

[@Explaining-AI]

You are most welcome :) With time this playlist will cover everything till yolov8 (and beyond) - kzbin.info/aero/PL8VDJoEXIjppNvOzocFbRciZBrtSMi81v For this video,I dont remember exactly, but I think somewhere around 2 weeks, including implementation, editing and everything.

[@EED-qj2tv]

how can we predict new images?

[@Explaining-AI]

Hi, For predicting new images, we would just need to call the FasterRCNN class's forward method with the image. github.com/explainingai-code/FasterRCNN-PyTorch/blob/main/tools/infer.py#L229 So the same call that is there in training Line 67 @01:05:46 , just that targets will be None

[@goneshivachandhra7470]

quality content in you channel Is detr covered in this series??

[@Explaining-AI]

Thank you! I think you missed my reply on your previous comment, so adding it here as well - yes it would cover DETR as well. After FasterRCNN, I plan to do Yolo/SSD/FPN and then I will get into DETR.

[@goneshivachandhra7470]

@@Explaining-AI Does detection series takes two months or more ?? I want to use detr in my project. I am finding difficult in understanding loss function of detr.

[@Explaining-AI]

@@goneshivachandhra7470 Yes unfortunately it would take time(not sure how much though) as I would do detr after covering yolo/ssd and few more. But since you need to use Detr now, I understand that this would be too late for you. If you think that it will be of some help to have a discussion on the loss function and the aspects you are finding difficult to understand, do send me an email regarding this and I can try to answer them and if needed, we can even get on a call for the same.

[@cryes9774]

i can understand just %50,i know whats happening here i listened previous explanation but i can't figure why did you do many things, do you think like that or im not capable for understand because of knowledge lackless

[@Explaining-AI]

Sorry about that. Though I did try my best to give explanations for the corresponding implementation, but if you are having trouble understanding then, clearly somethings could be improved. Could you tell me what specific parts of video you had doubt or trouble understanding in(or understanding the motivation). I will try to provide more explanations on them here.

[@cryes9774]

@@Explaining-AI your explanation is nearly perfect there is no need to improve i think because we re not child for explaine every single thing, i should to find connections between model explanation and fundtions but the topic is not easy i guess and im missing something,i will watch again and check the previous explanation just some times i m loosing my motivation and give up to struggle thank you ,i learnt a lot of things from your previous videos and i hope will

[@Explaining-AI]

@@cryes9774 Yes, FasterRCNN model indeed takes time to understand specially understanding its end to end implementation. When I was trying to understand what helped a lot was running a training iteration on a dataset and going through it step by step. Maybe you can try that as well. Just give it some time, and soon you will be able to get good grasp at it. And let me know if I can help in anyway.

[@daleanfer7449]

great❤

[@TamagoYakiLi]

Coding is very fast and hard to follow during the video. I need to keep pausing the video. but, anyway, thanks for your work

[@Explaining-AI]

Thank you for this feedback. Will try to decrease the speed a bit for future implementation videos. However that significantly increases the length of video, hence I try to achieve a balance between the two.