C4W3L01 Object Localization

Follow us:

[submagr]

As soon as I find a video by Prof Andrew on a topic I am looking for, I know this topic is so done for good.
Thanks, Prof for these wonderful lectures. I can't be enough grateful.

[dineshbh]

for object detection usually use 'blob ', contors to separate objects from background and classify that slits

[morsemo]

I am a total beginner(even for python).
I couldn't understand the courses here one month ago. Then I took about 1 month to go around and try most of the popular algorithms examples (with GPU linux server). Then I come back and watch the courses. Now I could be more confident to continue the course journey with Andrew.

[darinhitchings]

this is super good content thanks so much

[ElectricCircuitsLAB-ProfHazemA]

Thnaks a lot for useful and easy presentation

[constructor]

Since the loss function for the case when y1 = 0 includes only one term in contrast to the case when y1 = 0, isn't it kind of encouraging the network to predict that there is no object(background) over the other cases?

[RahulMahajanGoogle]

This is very good information & helpful. Thanks.

[harishkumaranandan]

Hi everyone I am confused about the bbox part. How does the feature vector stack in final FC layer spit out some arbitrary 4 number that are bbox parameters even before backpropagation and L2 loss part. Are these initial bbox co-ordinates the one of the feature vector that has the object in it? Lets say in the case of localizing a car the final FC layer before softmax will have learnt high level features like car wheels, windshield etc and at the end these are stacked. Having said the bbox of the whole car will be different than the bbox of high level features like wheel, windshield etc. I am confused in this part of predicting the initial bbox of the whole car even though it might not be accurate initially bathos does the bbox of high level feature vector match the bbox of whole object. correct me if i were wrong somewhere.

[computer-sci]

bounding box data as input, while training a model is give after convolution operation, am I right ?, I have little confusion . :)

[voovdhb]

In last part of this video, he said we can use softmax, squared error, logistic regression loss. if I use that, I think there will be three different type loss. And then how should I do back prop? Just calculate loss matched each output neuron's loss fucntion?

[AbhishekSinghSambyal]

Can you name any training set which has the same classes and bounding boxes values to try this approach?

[sandipansarkar]

nice explanation .need to watch agaian

[jasmineshaik]

If there is a pedestrian and car in a frame ??? Is it applicable

[saikrishnadyavarasetti]

If there 'n' of objects in an image, then how the softmax output will be? will be same [pc, bx, by, bw, bh, c1, c2, c3]? How the output will be?

[sudarsansudar]

What will be bx, by, bh, bw value in the output vector if multiple classes are present in the picture.

[priyankasn]

how we have given input image to each neuron

[praleen_]

Hi! I am wondering why the background is not included in the vectors!

[muhammedbuyukknac]

You can check out my repository over object localization for SINGLE object. It is a ready-to-run repository.

[guardrepresenter]

How pc variable know without calculating class labels.Because Andrew say if pc 0 the other variable dont care.But how pc know i am 0 or

[faridalijani]

how could one program "don't care" as an output of an image which contains no object?