Localization Object Detection
Localization
Find where the object is in the image and draw a bounding box around it.
Semantic Segmentation
Classify every pixel in the image to a class according to its context, so that each pixel is assigned to an object. Each colour is assigned to a type of object (in this case, red is cat, blue is dog, green is duck).
How can we achieve both?
Potentialy can use Fully Convolutional Networks. They are networks that only consist of convolutional layers without a need to be connected.
There were 2 problems with them:
FCNs couldn’t use any downsampling or pooling layers as it will mess up the location of the image instances
Many layers had to be added to learn both low-level and high-level features to maintain image resolution (computational expensive)
Solution: Encoders!
Encoders are typical CNNs such as ResNet that downsample. Decoders would be **deconvolutional **and up-sampling layers. This way, we can downsample to capture semantic information and recover the spatial information (both localization and semantic segmentation are achieved!).
Visual Comparison
However, another problem still persists. When the encoder downsamples, the image resolution is sacrificed. A way to solve this is using something called Skip Connections.
Skip Connections
Bypass layers and transfer the information intact to the next layers. i.e. Passes information from early layers to the encoder (skips the encoder).
Last updated
