If you want to spare some time: I won't tell the canonical truth here, so please do not read further if you are seeking for a definite answer in this topic. Most applications of convolutional networks just stick to one network architecture. Let us say, there is a biomedical image segmentation problem and people tend to use the so-called UNET for solving their problem. But why that one specifically? The answers to these questions usually do not satisfy a researcher like me: "because it has been applied successfully before to a range of problems", "because it is straightforward to implement", "because we tried this and it worked well". Even if these answers are not scientific enough, the points make sense and it seems that there is a certain randomness in how preferred architectures are selected for specific tasks.
However, it is important to note the differences between specific image segmentation tasks. Do we want to distinguish a large number of classes in complex RGB images or do we intend to segment large numbers of simple objects in greyscale images? Does track counting resemble distinguishing zebras from horses or lions from tigers or is it something more similar to counting the number of cars on a satellite image? Yes, you got that right and that is why it should not be entirely arbitrary what kind of network one uses for which specific task! Different network architectures are mainly developed because they fit some specific requirements in image analysis better than others. Luckily, some resources, such as the segmentation models website (smp.readthedocs.io/en/latest/index.html) provides some brief hints regarding the preferred network for certain tasks. Even more importantly, their code examples can be implemented such as the user can flexibly switch between a range of network architectures without the need to rewrite the ML script significantly!
0 Comments
Leave a Reply. |
FTAIGEThis project has received funding from the European |