C 5.2 | ConvNet Input Size Constraints | CNN | Object Detection | Machine learning | EvODN скачать в хорошем качестве

C 5.2 | ConvNet Input Size Constraints | CNN | Object Detection | Machine learning | EvODN

The problem we discussed in the previous video was that, using the Sliding window technique and taking the crop of the image at different locations, I am ending up with too many inputs to my Localization network. How do I reduce this? Since the ConvNets internally use the Sliding windows, we need not in fact crop the image at all the locations, we can feed the complete image as input to the network. So, this way, lets say, we will be able to take care of the Sliding Window issue? Now, what about the Image Pyramid with 6 different scales? This is where the problem comes. Since we know that the ConvNet expects inputs of fixed size, I will not be able to use the images of different scales as input. So, how do we solve this problem? To, solve this, lets first understand why the ConvNets expect fixed size input images in the first place. As, you will see in the video, the convolution and pooling layers have no problems processing images of different sizes. The restriction comes from the Fully Connected layer. Since the FC layer uses the dot product operation, it expects the Feature Vector to be of a fixed size. Now that we have understood the root cause of the problem, how do we fix this? To solve this, as we have already discussed in one of the earlier videos (   • C 4.11 | Fully Connected Layer as Conv Lay...  ), we can implement the FC layer as a Convolution operation. This will remove the fixed size restriction. But if you observe the output of this operation, you will see that, you will end up with different sized outputs. For a 6x6 image with a 3x3 convolution with Stride=1 and Padding=1 and 2x2 Pooling, the output will be a 3x3 matrix. With FC layer converted to Convolution operation, I have to use a 3x3 filter, which gives me a 1x1 output. Lets say, this will just say if there is an object or not. But if I scale this image (Image Pyramid) to 8x8, the output of the network will be 2x2. Now the question is, is this the output I am expecting? Does this make sense? If so, what does this 2x2 output mean, if I just want to know the presence or absence of an object? Lets discuss in the next video. ------------------------ This is a part of the course 'Evolution of Object Detection Networks'. See full playlist here:    • Evolution Of Object Detection Networks   ------------------------ Credits: http://host.robots.ox.ac.uk/pascal/VO... Copyright Disclaimer: Under section 107 of the Copyright Act 1976, allowance is made for “fair use” for purposes such as criticism, comment, news reporting, teaching, scholarship, education and research.