At 4:36 it is told that the lower the IOU threshold, the lower the precision. Isn't it true that the confidence threshold is meant here instead of the IOU threshold, because the lower the IOU threshold the faster something is considered a TP
from my understanding, if you have a lower IoU threshold, then you're more likely to mark your predictions positive (could be TP or FP). for lowest IoU for eg, if you mark everything positive, both your TP and FP increase, hence precision goes down. in other words, if you tell the model that hey you can mark it as positive even if the IoU is lesser than before, then the model wouldn't be strict as to how close it's predicted location is to the true location, and hence your predicted location's precision goes down
@@nb_bucky_my_beloved yes, it's an inherent trade-off! Decreasing IoU threshold, and thus increasing the number of detected objects (higher recall) will most likely result in more false positives (lower precision).
It's technically the same thing depending on the object detection technique we're talking about. In earlier YOLO versions, the predicted confidence is compared to a target confidence (which is the computed IoU)
If various IOU Thresholds produce a PR-Curve then what is the significance of classification various thresholds? Is classification threshold (that controls the TP, TN, FP and FN) used anywhere?
It's the other way around! Decreasing IoU threshold, results in less objects being left out from being detected, thus the higher recall. But also, more objects will be wrongly detected, thus the lower precision.
CV can never solution the gun Detection problem in school, because ML suffer with adversary attacks. Only Right education and Spritual knowledge in school will solve human problems as it work directly on human intelligence. Yours video was good and knowledgeable 👍👌
Confusion Matrix never made easy before .. the example of gun was really cleaver .
This video deserves more likes
Great comic! finally this confusion matrix is no more confusion!
I watched this to relax, endee up understanding the whole concept... well done sir
I'm glad you were able to understand 😁
At 4:36 it is told that the lower the IOU threshold, the lower the precision. Isn't it true that the confidence threshold is meant here instead of the IOU threshold, because the lower the IOU threshold the faster something is considered a TP
yes, also my understanding! I guess this video informs wrongly
from my understanding, if you have a lower IoU threshold, then you're more likely to mark your predictions positive (could be TP or FP). for lowest IoU for eg, if you mark everything positive, both your TP and FP increase, hence precision goes down.
in other words, if you tell the model that hey you can mark it as positive even if the IoU is lesser than before, then the model wouldn't be strict as to how close it's predicted location is to the true location, and hence your predicted location's precision goes down
@@nb_bucky_my_beloved yes, it's an inherent trade-off! Decreasing IoU threshold, and thus increasing the number of detected objects (higher recall) will most likely result in more false positives (lower precision).
It's technically the same thing depending on the object detection technique we're talking about.
In earlier YOLO versions, the predicted confidence is compared to a target confidence (which is the computed IoU)
great explanation. Tks for your afford!!!
This really helped a lot. Thanks 🙋♂️
amazing thanks a lot
Isn't the PR curve plotted by varying the SCORE/CONF threshold? (not the IOU threshold, as stated in the video)
Ok thanks I understand, so why don’t we have AR average recall?
If various IOU Thresholds produce a PR-Curve then what is the significance of classification various thresholds? Is classification threshold (that controls the TP, TN, FP and FN) used anywhere?
Does payment accept paypal?
Yes it does
Thank you very much. Why please The higher the IoU the higher the recall:
It's the other way around! Decreasing IoU threshold, results in less objects being left out from being detected, thus the higher recall. But also, more objects will be wrongly detected, thus the lower precision.
CV can never solution the gun Detection problem in school, because ML suffer with adversary attacks.
Only Right education and Spritual knowledge in school will solve human problems as it work directly on human intelligence.
Yours video was good and knowledgeable 👍👌
At the least, we can spread awareness. It may not be a solutions but we can work towards early warning systems.