Classification and Segmentation
Part 1: Image Classification
In this part we use FashionMNIST as the input dataset to train a classification model. Here are some samples from the dataset.
The architecture of my model is conv->relu->maxpool->conv->relu->maxpool->fc->relu->fc.
Here are the accuracy curves of my model.
Here’s the per class accuracy of my classifier. We can see that Shirts are the hardest to get.
class | correct | wrong |
---|---|---|
T-shirt | ||
Trouser | ||
Pullover | ||
Dress | ||
Coat | ||
Sandal | ||
Shirt | ||
Sneaker | ||
Bag | ||
Ankle boot |
Here are the visualization of the filters from the first layer of convolution.
Part 2: Semantic Segmentation
The architecture of my model is :
lr=1e-3, weight_decay=1e-5, batch_size=1, epoch=30
And the accuracy curve is:
The average AP is:
Here are some of the result of my collection(the input have been reshape in to 256*256):
original | output |
---|---|
We can see that facade(blue) and window(orange) is the easiest to obtain while pillar(green) and balcony(red) is the hardess to get.