giovana-morais / color_cluster Goto Github PK

atualmente, os arquivos são salvos apenas como nome_do_arquivo.pkl, independente do número de cores que foram extraídas dele.

o problema disso é que o método de clusterização (kmeans ou vector_quantization) pode se confundir ao encontrar um arquivo com nome correto, mas número de cores diferentes das que recebeu como parâmetro.

pra corrigir isso provavelmente é só passar o número de cores pro nome do arquivo, por exemplo, nome_arquivo_3.pkl para apenas três clusters.

when a .pickle file is in the same folder that files with image extensions, like .jpg, .jpeg etc etc, the program crashes because it tries to apply image functions to those files.

maybe save pickle files in another folder or just try to ignore them when reading the images.

after getting every image pallette, we need to compare it and check how similar they are, so we can group it.

nowadays, we have a method get_image_clusters that should work, but it doesn't because of the clusters' dimension. i'm not sure if we can reuse the method or if we'll have to implement another one from scratch, but this grouping images method must be implemented.

major steps

1. refactor get_colors
2. order palettes so comparison may be easier (issue #7)
3. refactor/rewrite get_image_clusters

nowadays, the pallette generated doesn't has any order and I *personally * think that the visualization may be nicer if we order it from most used color to less used.

it would be nice to standardize the code use a formatter like black. flake8 may also be a option to have a nicer and readable code.

Pillow has a quantize method that changes image to use only the n dominant colors. despite the fact that I believe this method ALSO uses KMeans under it, I think it will be nice to try to use it too and then compare performances.

it must contain a detailed explanation about the following cases:

1. argument is a folder

• if this happens, images should be grouped by their similarity and saved in different folders to represent the clusters. or maybe just output a list with the clusters participants.

2. argument is a file

• if only a file is provided, code should return only the clusters of that image.

3. a file and a folder are passed

• basically, the program should get the image and return the images related ONLY to that image.

Usar o KMeans para fazer o agrupamento de pixels e encontar os n grupos de cores mais significativos

apart from KMeans, there are a ton of different clustering methods and it would be nice to compare the overall performance.