Great work. I like how you made youtube chapters to explain independent techniques like NMS. Really useful. Many people don't have the time to go through papers in details and just run the codes to get things done. Your videos could be helpful to solve that problem. I'm personally hoping to see videos on YOLO series especially the YOLOX model :) You could also talk about the object detection models landscape and how each model has pros/cons w.r.t. inference time (FPS) and performance.
Nice, this topic deserves its own playlist. RCNN has so many component, you can make separated short video for each component, so it wont be overwhelming for the viewers.
Thanks, Muhammad. I actually want to create videos for other object detection algorithms as well and put them in a playlist. From my past experience and based on the videos I've seen, usually, long videos get more viewers. I already separated this video into different chapters and viewers can watch each one on their own time. It's a kinda subjective opinion I believe.
@@soroushmehraban Oh I spoke too fast, (bc I watched some parts of the video several times, I thought you used the expression several times)... Yeah I take it back apologies, oc everyone can use this expression!
Very Nice Explanation Just one question they use SVM in the final steps to make prediction. Is that for the class prediction or for the Bounding Box prediction. Also how to we know that we need to predict (c + 1) classes ? Do we know beforehand what classes of objects are present in our image ?
Thanks, Mohandes. I'll try enhancing the quality by changing my recording method but still it's not gonna be perfect. At least not in the first few videos.
That's a great question. I think I should have mentioned that. Our model might predict different bounding boxes pointing to the same object. In such a scenario, we do the following: 1) Sort all the predicted bounding boxes based on the class score (In descending order). 2) Pick the first bounding box that has the highest probability score. 3) Compute the IoU of the selected bounding box with other bounding boxes pointing to the same class. 4) If the IoU of any bounding box with this bounding box is larger than a threshold (such as 0.5), then we remove the bounding box having the lower class score. I hope it's clear.
@@soroushmehraban i think following conditions might not be sufficient, because even if we sort and pick highest one... again we left with question of all these are pointing to same object location or reference really in a image ? same object references might be at multiple places please clarify this doubt
That's true we might have same objects at multiple places. let's say we have object A at location (x1, y1) and (x2, y2). for location (x1, y1) our model might predict multiple bounding boxes all refer to the object A. Out of all these bounding boxes we only keep the one that has the highest score and others if they have IOU higher than a threshold with this bounding box, we remove them. For object A at place (x2, y2), since it's in different area of the image, the IoU with the one having highest score is less than a threshold, so we keep the second one having the highest threshold and again others having IoU higher than a threshold, we remove them. @@NagarajuSeru-rc7lb
Very Nice.. Thank you so much.... I have a question related to NMS... that As you explained about NMS, IOU of classified object regions will calculated over the ground truth value at the time of training and validation but what about at the time of inference ? since you have grouth truth values at time of train and validate only but not at inference. awaiting for your response.... thank you so much adavance
![The moment we stopped understanding AI [AlexNet]](http://i.ytimg.com/vi/UZDiGooFs54/mqdefault.jpg)
One of the best videos I have watched. Very detailed Explanations. Keep up the good work
Thanks 🙂
Great Work! You explained 1000 times better than my uni lecturer :D
bro you did actually the best video for eexpaling Rcnn
Such a great video!! Keep them coming!
Such an underrated video. Well done mate!
Glad you enjoyed it!
the way you organised the following content are just awesome ..
Great work.
I like how you made youtube chapters to explain independent techniques like NMS. Really useful.
Many people don't have the time to go through papers in details and just run the codes to get things done.
Your videos could be helpful to solve that problem.
I'm personally hoping to see videos on YOLO series especially the YOLOX model :)
You could also talk about the object detection models landscape and how each model has pros/cons w.r.t. inference time (FPS) and performance.
Wonderful feedback, Gota. I'll make sure to create them in the future
Very nice! I can't wait to see more videos like this!
Thanks, Jeffrey! Wait for the better ones then 😄
oh my it explains everything at once! Thank you for making this video!
This so easy how i can uderstand about RCNN and that is because your explanation!
thank you very much, i love your video
Glad you liked it!
Simple and easy to understand! Thank you for making this video :)
Glad it was helpful!
Thanks for your work! It's helps me a lot! Appreciate that~
Very well explained . Thank you
Thanks a lot for this! It was really clean and precisely explained. mAP explanation was on point.
Glad you liked it!
Nice, this topic deserves its own playlist. RCNN has so many component, you can make separated short video for each component, so it wont be overwhelming for the viewers.
Thanks, Muhammad. I actually want to create videos for other object detection algorithms as well and put them in a playlist. From my past experience and based on the videos I've seen, usually, long videos get more viewers. I already separated this video into different chapters and viewers can watch each one on their own time. It's a kinda subjective opinion I believe.
@@soroushmehraban how about Yolo?
This is great. Nice work!! Waiting for more such videos.
Thanks, Raghuveer! Appreciate it.
Very nicely explained with animation 💜
Awesome video Now I can read the paper and use the video as a guide.
Glad you liked it!
Cool video! Keep them coming
Thanks, Mohamed!
Cool! Nice work.
Thanks, Seokeon. I hope you find it useful.
Cool! Nice work💥
Good job Soroush, Very nice video! It helped me a lot specially to understand the mAP metric. Just Keep going :)
Glad you liked it :)
Nice video! Keep up the great work
Thank you, Bellz!
best explanation ever!
I really appreciate it, very good explanation. Thanks!
clean explanation give this man more sub !
Well done. That was great
Thanks Aref
Thanks very much for this, it's much clearer to me know (after starting from just the paper). (Edit : this Paper is clearly explained in every way)
Thanks for the honest feedback 😃 looking at the previous videos posted, I’m not using that phrase anymore.
@@soroushmehraban Oh I spoke too fast, (bc I watched some parts of the video several times, I thought you used the expression several times)... Yeah I take it back apologies, oc everyone can use this expression!
Very interesting! need more videos.
dude!!! that was such a nice explanation
Thanks!
Great video. Good job. Request for follow up videos: Faster R-CNN, Mask R-CNN, DETR
Thanks, Yaser. I'll post them. But first I'll post Fast R-CNN
Informative video!
Thanks, Kaan!
great work!
thank you for your great explanation! keep going!
Thanks!
Great video. keep up the good work
Great explanation, keep doing it!
Thanks, Alexander!
Great explanation
Nice job! Keep up the good work!
Thanks for the positive energy, Chayan!
Congrats. Good work.
Thanks, Can! Appreciate it.
Great explanation❤
Nice one! Please make more
Thanks, Ishaan. Sure!
Very Nice Explanation
Just one question they use SVM in the final steps to make prediction. Is that for the class prediction or for the Bounding Box prediction.
Also how to we know that we need to predict (c + 1) classes ? Do we know beforehand what classes of objects are present in our image ?
bright explanation Thanks
Thanks, Alireza. I hope you found it useful.
literally , Clearly EXPLAINED
Nice work
Thanks, Pouya.
Keep up the good work
Thanks!
good work
Thanks, Lakshay.
Thank you so much
Very nice! Thanks a lot! May you please upload your slides, too?
داداش دمت گرم
Great Job, Can't wait to see more videos of you. Can you fix your microphone for next videos?
Thanks, Mohandes. I'll try enhancing the quality by changing my recording method but still it's not gonna be perfect. At least not in the first few videos.
Great.
nice one thx
Thank u
thank you so much , such an amazing video . Can i ask which tool/app you using for this slide? i love how they working
Thanks for the feedback Huy 🙂It's just a powerpoint.
what is the background music you are using in the video ?
I don't remember that was a long time ago. I'm not using any background music anymore.
great
Thanks, Louis.
How does NMS works in inference? As we won't be having ground truth
That's a great question. I think I should have mentioned that. Our model might predict different bounding boxes pointing to the same object. In such a scenario, we do the following:
1) Sort all the predicted bounding boxes based on the class score (In descending order).
2) Pick the first bounding box that has the highest probability score.
3) Compute the IoU of the selected bounding box with other bounding boxes pointing to the same class.
4) If the IoU of any bounding box with this bounding box is larger than a threshold (such as 0.5), then we remove the bounding box having the lower class score.
I hope it's clear.
@@soroushmehraban i think following conditions might not be sufficient, because even if we sort and pick highest one... again we left with question of all these are pointing to same object location or reference really in a image ? same object references might be at multiple places
please clarify this doubt
That's true we might have same objects at multiple places. let's say we have object A at location (x1, y1) and (x2, y2). for location (x1, y1) our model might predict multiple bounding boxes all refer to the object A. Out of all these bounding boxes we only keep the one that has the highest score and others if they have IOU higher than a threshold with this bounding box, we remove them. For object A at place (x2, y2), since it's in different area of the image, the IoU with the one having highest score is less than a threshold, so we keep the second one having the highest threshold and again others having IoU higher than a threshold, we remove them. @@NagarajuSeru-rc7lb
Very Nice.. Thank you so much....
I have a question related to NMS... that
As you explained about NMS, IOU of classified object regions will calculated over the ground truth value at the time of training and validation but what about at the time of inference ? since you have grouth truth values at time of train and validate only but not at inference.
awaiting for your response.... thank you so much adavance
Really good tut, but background music is disturbing attenttion.
@@AntonMorzhakov Totally Agree. Removed it from next videos 👍
Kara Road
Bauch Estates
Fudge, you copy other's work
Nais work man, keep this up, I wanna see moo 🤌❤️
Thanks, man! I'll try my best.