Finnaly someone explaining how they actually make detections. So many people on here gloss over that part, and explain nothing about how the featuremaps are turned into boxes.
Thank you for this wonderful Video, May I know how the default bounding boxes are encoded and what are their default location , size and aspect ratio in a feature map ?
Where can I get coordinates of the default 8732 boxes? I need to interpret the raw output localization tensor (the one with size [1,4,8732]) using C++ (LibTorch) so I can't use nvidia_ssd_processing_utils for that. I understand how to calculate the final bounding boxes but I don't have the default box data (dx, dy, dw, dh).
Beautiful lecture, thank you sir! May I just ask a little question? On the GIF visualizations that you show, it intuitively seems like we perform convolution of default bounding boxes with the feature maps of different scales. But it is written, that the actual size of convolutional blocks is 3 x 3. Could you clarify, how convolutions of feature maps with 3 x 3 blocks "intuite", with which default bounding box the network is working? Thank you very much! Sincerely, Pavel
in 6:27 you have 6 results concatenated in the detection block (detection 8732 by class), but in 11:55 you have only 5 outputs (features maps) ??????????????????????????????????????????????????????????????????????????????????????????????!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
amasing video , clear and direct. please never stop making these !! thank you
Amazing lecture. Very good explanation of SSD and its EfficientDet descendants.
Best description I found. Good presentation. I love the visuals.
You have fantastic pedagogical skills. Thanks for the clear and lovely explanations
Many compliments. This is a great introduction to SSD and its directions for improvement.
Finnaly someone explaining how they actually make detections. So many people on here gloss over that part, and explain nothing about how the featuremaps are turned into boxes.
Very clear explanation and description of basically the original paper. Thank you!
This is really ground breaking! I love your explanation. Straight to the point & Simple.
We'll pass the feedback on! :)
Thank you for this wonderful Video, May I know how the default bounding boxes are encoded and what are their default location , size and aspect ratio in a feature map ?
This is so clean and thorough!
best explaination on ssd i have ever found!
Where can I get coordinates of the default 8732 boxes?
I need to interpret the raw output localization tensor (the one with size [1,4,8732]) using C++ (LibTorch) so I can't use nvidia_ssd_processing_utils for that. I understand how to calculate the final bounding boxes but I don't have the default box data (dx, dy, dw, dh).
Thanks for your generosity to teach this topic.
Could you please make videos on RCNN and its variants with focus on RPN?
Thanks a lot, great and clear explanation
Beautiful lecture, thank you sir! May I just ask a little question?
On the GIF visualizations that you show, it intuitively seems like we perform convolution of default bounding boxes with the feature maps of different scales. But it is written, that the actual size of convolutional blocks is 3 x 3. Could you clarify, how convolutions of feature maps with 3 x 3 blocks "intuite", with which default bounding box the network is working?
Thank you very much! Sincerely, Pavel
very clear and helpful thank you
Can I use your presentation for my university project?
You’re wonderful and save my life!!!!
it was very useful and thank you for your time!
such a great video, thank you!
Thanks for sharing, good explanation.
Great lecture.
crystal clear explaination
This was amazing!
Very good explanation, but it seems your layers are not correct they are 6 layers not 5.
is ssd using FFNN or CNN ?
I like the jib of this guy’s jab
in 6:27 you have 6 results concatenated in the detection block (detection 8732 by class), but in 11:55 you have only 5 outputs (features maps) ??????????????????????????????????????????????????????????????????????????????????????????????!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
svm = __support__ vector machine.
I kept thinking that myself lol