I am not in this competition yet (and I'm afraid I won't), so I would like to share my idea of such an image processing technique - or just a feature.
\nLooking at these images one can realize, that these animals are just like graphs, and has some very interesting features in it. For example a round with 4 "roads" into it. I'm wondering about a solution that could standardize these pictures:
\n1. adjacency matrix (based on the edges of images as the vertex of the graph, and the neighbourhood as the edge of the graph) and computing the features, like number of triangles, etc.
\n2. edge distribution in categories
\n3. Canonical form of the graph to simply check the similarity
\n\nThere are some other graph similarity projects, which may also help in categorization.
\n\nYou know I hope that understanding the problem sometimes gives a stronger solution than a simple brute force :) So please someone try this out for me, because I won't have time for it, but I'm keen to know the effectiveness. So a response about the improvement would be nice too :D
\nsome useful links:
\nhttp://politespider.com/papers/general/Graph%20similarity%20algorithms.pdf
https://www.cs.cmu.edu/~jingx/docs/DBreport.pdf
https://wadsashika.wordpress.com/2014/09/19/measuring-graph-similarity-using-neighbor-matching/
http://docs.lib.purdue.edu/cgi/viewcontent.cgi?article=2761&context=cstech
http://i11www.iti.uni-karlsruhe.de/_media/projects/spp1126/files/bl-ssg-04.pdf
I am not in this competition yet (and I'm afraid I won't), so I would like to share my idea of such an image processing technique - or just a feature.
\nLooking at these images one can realize, that these animals are just like graphs, and has some very interesting features in it. For example a round with 4 "roads" into it. I'm wondering about a solution that could standardize these pictures:
\n1. adjacency matrix (based on the edges of images as the vertex of the graph, and the neighbourhood as the edge of the graph) and computing the features, like number of triangles, etc.
\n2. edge distribution in categories
\n3. Canonical form of the graph to simply check the similarity
\n\nThere are some other graph similarity projects, which may also help in categorization.
\n\nYou know I hope that understanding the problem sometimes gives a stronger solution than a simple brute force :) So please someone try this out for me, because I won't have time for it, but I'm keen to know the effectiveness. So a response about the improvement would be nice too :D
\nsome useful links:
\nhttp://politespider.com/papers/general/Graph%20similarity%20algorithms.pdf
https://www.cs.cmu.edu/~jingx/docs/DBreport.pdf
https://wadsashika.wordpress.com/2014/09/19/measuring-graph-similarity-using-neighbor-matching/
http://docs.lib.purdue.edu/cgi/viewcontent.cgi?article=2761&context=cstech
http://i11www.iti.uni-karlsruhe.de/_media/projects/spp1126/files/bl-ssg-04.pdf
>So please someone try this out for me, because I won't have time for it, but I'm keen to know the effectiveness.
\n
If you provide some implementation ) And to check effectiveness you don`t need to get results until deadline $)
alright, it seems that the community don't like theoritical questions, so here is a (bit mistaken, but more-or-less correct) implementation of the graph (it works on the solution of https://www.kaggle.com/c/datasciencebowl/details/tutorial):
\n\nimport networkx as nx
G=nx.Graph()
i = 1
for y in range(0, im[1].size-2):
for x in range(0, im.size/im[1].size-2):
if labels[x, y] == 3:
G.add_node(x*(y+1) + (y+1))
i = i+1
if labels[x, y-1] == 3:
G.add_edge(x*(y+1) + (y+1),x*y + y)
if labels[x-1, y] == 3:
G.add_edge(x*(y+1) + (y+1),(x-1)*(y+1) + (y+1))
if labels[x-1, y-1] == 3:
G.add_edge(x*(y+1) + (y+1),(x-1)*y + y)
if labels[x, y+1] == 3:
G.add_edge(x*(y+1) + (y+1),(x+1)*y + y)
G=nx.Graph(G)
\n\nG has a degree histogram (nx.degree_histogram(G)) of :
\n[0,
212,
247,
167,
41,
197,
236,
243,
1556,
47,
19,
19,
4,
25,
33,
19,
51,
2,
3,
2,
0,
1,
3,
1,
1,
0,
0,
0,
0,
1]
which shows, that I started Python today, but also that it works almost well (the maximum would be 8 in theory).
", "rawMarkdown": "", "votes": null } ], "comments": [ { "id": 64934, "author_name": "oldufo", "author_url": "", "post_date": "02/26/2015 11:25:51", "content": ">So please someone try this out for me, because I won't have time for it, but I'm keen to know the effectiveness.
\n
If you provide some implementation ) And to check effectiveness you don`t need to get results until deadline $)
alright, it seems that the community don't like theoritical questions, so here is a (bit mistaken, but more-or-less correct) implementation of the graph (it works on the solution of https://www.kaggle.com/c/datasciencebowl/details/tutorial):
\n\nimport networkx as nx
G=nx.Graph()
i = 1
for y in range(0, im[1].size-2):
for x in range(0, im.size/im[1].size-2):
if labels[x, y] == 3:
G.add_node(x*(y+1) + (y+1))
i = i+1
if labels[x, y-1] == 3:
G.add_edge(x*(y+1) + (y+1),x*y + y)
if labels[x-1, y] == 3:
G.add_edge(x*(y+1) + (y+1),(x-1)*(y+1) + (y+1))
if labels[x-1, y-1] == 3:
G.add_edge(x*(y+1) + (y+1),(x-1)*y + y)
if labels[x, y+1] == 3:
G.add_edge(x*(y+1) + (y+1),(x+1)*y + y)
G=nx.Graph(G)
\n\nG has a degree histogram (nx.degree_histogram(G)) of :
\n[0,
212,
247,
167,
41,
197,
236,
243,
1556,
47,
19,
19,
4,
25,
33,
19,
51,
2,
3,
2,
0,
1,
3,
1,
1,
0,
0,
0,
0,
1]
which shows, that I started Python today, but also that it works almost well (the maximum would be 8 in theory).
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "64927": "", "64934": "", "64959": "" }, "source": "meta" }