I almost finished 1/3 of your uploaded videos. It feels like someone is reading papers with me. The feeling is great! Thanks so much Yannic. Keep it up!
This shows that there still so much left in DL to be done. One thing I see is that it seems that every point is making a prediction. According to research by Uber, there are location sensitive CNNs which can also be tried in these cases. Would love to see something like a combination of the two ideas.
These embeddings get more interesting the more you think about it. It is essentially two neural networks inventing a language to talk to each other. If we can make this interpretable, it might open up a lot of possibilities.
Corner pooling is a really smart way to largely increase perception field, sort of like deformable convolution. But DETR seems to generally solve the problem in object detection since it make use of the full image as perception field.
Any specific reason you went with this approach over any of the other very similar boxless/keypoint detection approaches like CenterNet ("Objects as Points") or CSPNet ("Center and Scale Prediction") that not even require laborious embeddings while performing equally or even better? Or the "CenterNet: Keypoint Triplets for Object Detection" paper that basically is the combination of the CornerNet and the Center approaches. I mean they basically all do the same (keypoint detection) which in my opinion is quite different to what you suggested with the cross attention matrix from the attention heads?
My intuition is that using paired keypoints is cheaper but should be more inaccurate over anchor boxes. For example, It is not clear what the paper does when there are overlapping objects that share the same keypoint (e.g. top-left). Using keypoints is interesting nevertheless. I found another recent paper that just uses keypoints inside transformer to replace RGB tracking and matching pipeline for pose tracking task: arxiv.org/pdf/1912.02323.pdf
Your videos are great!! Keep them up :) Why do you think they decided to go with these push and pull losses instead of using a triplet loss? Seems almost identical to the push + pull losses they propose
I almost finished 1/3 of your uploaded videos. It feels like someone is reading papers with me. The feeling is great! Thanks so much Yannic. Keep it up!
Paper from 1 year ago is now "a bit old". Just amazing how fast the field moves.
This shows that there still so much left in DL to be done. One thing I see is that it seems that every point is making a prediction. According to research by Uber, there are location sensitive CNNs which can also be tried in these cases. Would love to see something like a combination of the two ideas.
Thank you for the explanation.
These embeddings get more interesting the more you think about it. It is essentially two neural networks inventing a language to talk to each other. If we can make this interpretable, it might open up a lot of possibilities.
first subscription in my life. thanks for your video
Corner pooling is a really smart way to largely increase perception field, sort of like deformable convolution. But DETR seems to generally solve the problem in object detection since it make use of the full image as perception field.
Nicely explained
I really like your content, can you make an explanation of centernet : objects as point i dont really quite get the idea of its loss function
Any specific reason you went with this approach over any of the other very similar boxless/keypoint detection approaches like CenterNet ("Objects as Points") or CSPNet ("Center and Scale Prediction") that not even require laborious embeddings while performing equally or even better? Or the "CenterNet: Keypoint Triplets for Object Detection" paper that basically is the combination of the CornerNet and the Center approaches.
I mean they basically all do the same (keypoint detection) which in my opinion is quite different to what you suggested with the cross attention matrix from the attention heads?
Yes this paper didn't turn out to be exactly what I hoped, but still interesting. I chose it just because it sounded like fun.
The person who put the thumbs down has Oppositional defiant disorder (ODD)🤣
My intuition is that using paired keypoints is cheaper but should be more inaccurate over anchor boxes. For example, It is not clear what the paper does when there are overlapping objects that share the same keypoint (e.g. top-left). Using keypoints is interesting nevertheless. I found another recent paper that just uses keypoints inside transformer to replace RGB tracking and matching pipeline for pose tracking task: arxiv.org/pdf/1912.02323.pdf
Hey, have you done videos on the older but still heavily used architectures Faster RCNN, SSD, YOLO3, RetinaNet?
How can i find the research paper like you do
So, does network predict a tensor of WxHxC for heatmap branch ?
Yes, one for top left and one for bottom right
Your videos are great!! Keep them up :) Why do you think they decided to go with these push and pull losses instead of using a triplet loss? Seems almost identical to the push + pull losses they propose
No idea, but it's either the first thing they tried, or they tried a bunch of things and this worked the best.
now Taiwan (GMT+8) is 11 PM
yt: it's time reading a paper
embeddings of 1 dimension, not 1 number. 1 number wouldn't work lol
first