PROBLEM OF UNDERSTANDING ? if you don't understand why at 0:28, Andrew Ng introduces a grid that seems to come out of nowhere, or if you don't understand well the logical sequence of this week's videos, it is simply because the video C4W3L05 has been forgotten in this playlist ! Everything will be clearer if you look at that missing video Regards
So, the main idea is that, first: find the box which has the highest Pc value, no matter where that is in the image. Second, get the iou values of the boxes nearby and kill the boxes that are overlaped with the chosen box a lot (chosen due to the highest Pc as I mentioned) and repeat that with the boxes survived and one which has the second highest Px..
6:02 using IOU intuitively i think this helps in detecting close by objects of same type as we only discard the boxes that are trying to bound the same object leaving the one that is bounding a different one
What's the difference between the output box (bx,by,bh,bw) and anchor box mentioned in the previous lecture. Anchor box is predefined, so, here, for each grid cell, it outputs an anchor box? I am super confused about this. Besides, if we run a classification algorithm on each grid cell, how even possible the output box is bigger than the grid cell? Thank you!
If I understood correctly, for each grid cell, YOLO outputs pre-specified number of anchor boxes, say 5 anchor boxes. And each of the anchor box can be a bounding box prediction. So, for example if you have 3 x 3 grid cells and 5 anchor boxes, in total YOLO proposes 3 x 3 x 5 = 45 bounding boxes. Most bounding boxes found through anchor boxes are useless because they are overwrapping to each other (hence we need Nonmax suppression). The number of anchor boxes is predefined. The "specialization" of anchor boxes is also predefeind. (e.g. 1st anchor box tends to capture small bounding box, 2nd anchor box is specialized to capture narrow objects with narrow bounding box... the 5th anchor box is specialized to capture large objects with large bounding box). I tried to study the details of YOLO and wrote my learning progress at: fairyonice.github.io/tag/object-detection-using-yolov2-on-pascal-voc2012-series.html Nonmax suppression is particularly discussed at fairyonice.github.io/Part_6_Object_Detection_with_Yolo_using_VOC_2012_data_inference_image.html
Well, Why dont we just pick the bounding box with highest pc, instead of doing the second step, where we discard any other box intersecting with highest pc? seems unnecessary
No, he means to discard bounding boxes with high IOU. If there is a high IOU between 2 boxes, it is likely that they are trying to bound the same object, resulting in a unwanted duplicate box. So, discard one. (Conversely, if the IOU is low, then there is a higher chance that the boxes are for different objects since there is less overlap.) It confused me at first too.
PROBLEM OF UNDERSTANDING ?
if you don't understand why at 0:28, Andrew Ng introduces a grid that seems to come out of nowhere, or if you don't understand well the logical sequence of this week's videos, it is simply because the video C4W3L05 has been forgotten in this playlist !
Everything will be clearer if you look at that missing video
Regards
I found the missing video - czcams.com/video/gKreZOUi-O0/video.html . Best Of Luck
THANK YOU VERY MUCH!!!
Thank you for this!
But this video is not available in youtube 😢
great explanation, thank you, sir!
So, the main idea is that, first: find the box which has the highest Pc value, no matter where that is in the image. Second, get the iou values of the boxes nearby and kill the boxes that are overlaped with the chosen box a lot (chosen due to the highest Pc as I mentioned) and repeat that with the boxes survived and one which has the second highest Px..
6:02 using IOU intuitively i think this helps in detecting close by objects of same type as we only discard the boxes that are trying to bound the same object leaving the one that is bounding a different one
As you said more IoU value gives a more accurate bounding box, but in this video you are discarding the boxes if IoU>=0.5 , why?
For me, this was one of the most difficult explanations I have ever heard.
What's the difference between the output box (bx,by,bh,bw) and anchor box mentioned in the previous lecture. Anchor box is predefined, so, here, for each grid cell, it outputs an anchor box? I am super confused about this. Besides, if we run a classification algorithm on each grid cell, how even possible the output box is bigger than the grid cell? Thank you!
If I understood correctly, for each grid cell, YOLO outputs pre-specified number of anchor boxes, say 5 anchor boxes. And each of the anchor box can be a bounding box prediction. So, for example if you have 3 x 3 grid cells and 5 anchor boxes, in total YOLO proposes 3 x 3 x 5 = 45 bounding boxes. Most bounding boxes found through anchor boxes are useless because they are overwrapping to each other (hence we need Nonmax suppression).
The number of anchor boxes is predefined. The "specialization" of anchor boxes is also predefeind. (e.g. 1st anchor box tends to capture small bounding box, 2nd anchor box is specialized to capture narrow objects with narrow bounding box... the 5th anchor box is specialized to capture large objects with large bounding box). I tried to study the details of YOLO and wrote my learning progress at: fairyonice.github.io/tag/object-detection-using-yolov2-on-pascal-voc2012-series.html
Nonmax suppression is particularly discussed at fairyonice.github.io/Part_6_Object_Detection_with_Yolo_using_VOC_2012_data_inference_image.html
It only tells where is the enter, the bounding box is specified with the 4 numbers
very nice expression
Sir ,what is the need of this NMS, why cant we take the box diretly with high probability score.
does this mean that is it impossible to predict 2 cars behind each other?
My hunch is this only happens at the inference time. Right?
I believe so.
Well, Why dont we just pick the bounding box with highest pc, instead of doing the second step, where we discard any other box intersecting with highest pc? seems unnecessary
Because you can have an object appearing in multiple places.
3:23 Get rid of anything with low Iou and high confidence he meant ?
No, he means to discard bounding boxes with high IOU. If there is a high IOU between 2 boxes, it is likely that they are trying to bound the same object, resulting in a unwanted duplicate box. So, discard one. (Conversely, if the IOU is low, then there is a higher chance that the boxes are for different objects since there is less overlap.) It confused me at first too.
@@srishtikumar5544 nice one!
@@srishtikumar5544 thank you
Notification alert: cash is on its way back