In [1]:

    

from __future__ import division, print_function


    

In [2]:

    

%matplotlib inline
import matplotlib.pyplot as plt
import seaborn as sns


    

In [3]:

    

import numpy as np


    

In [4]:

    

import sys
sys.path.append('/zfsauton/home/dsutherl/scratch/caffe/python')
import caffe


    

In [5]:

    

from IPython.display import HTML, display, Image


    

In [6]:

    

import os


    

In [7]:

    

# img = os.path.join('yemen_odo_user_images', os.listdir('yemen_odo_user_images/')[3])
img = 'flag2.jpg'


    

In [8]:

    

with open('format.html') as f:
    display(HTML(f.read().replace('__IMAGE_PATH_HERE__', 'flag2.jpg')))


    

    







  

    
    

    
    
Coords
  

    

In [9]:

    

# flag cutout
coords = [[262,3],[294,34],[337,45],[394,76],[380,248],[367,287],[381,346],[303,267],[268,206],[264,167],[217,130],[180,107],[173,80],[201,15],[208,0],]


    

In [10]:

    

import random
with open('all_images') as f:
    filenames = [s.strip() for s in f]
random.shuffle(filenames)


    

In [11]:

    

import cv2


    

In [12]:

    

backgrounds = []

for l in filenames:
    dst = cv2.imread(l)

    #discard images with extreme height to width ratio
    yd, xd = dst.shape[:2]
    dratio = yd / xd
    if dratio > 5.0 or dratio < 0.2:
        print('{}: bad ratio {}'.format(l, dratio))
        continue

    backgrounds.append(dst)
    if len(backgrounds) >= 5000:
        break


    

In [13]:

    

from compositing import generate_composites


    

In [14]:

    

%%time
composites = list(generate_composites(img, backgrounds[:1000], np.array(coords),
                                      ))#randomize_loc=True, rand_scale=True))


    

    




CPU times: user 980 ms, sys: 375 ms, total: 1.35 s
Wall time: 1.35 s

In [15]:

    

plt.imshow(cv2.cvtColor(composites[0], cv2.COLOR_BGR2RGB))
plt.grid(False)


    

In [16]:

    

from extract_features import get_features, load_net


    

In [17]:

    

%%time
layers = ['fc6', 'fc7']
batch_size = 2000
net, transformer = load_net(layers=layers, batch_size=batch_size)


    

    




CPU times: user 687 ms, sys: 801 ms, total: 1.49 s
Wall time: 1.52 s

In [18]:

    

%%time
embeddings = list(get_features(
    [(c / 255)[:, :, (2,1,0)] for c in composites],
    layers=layers, batch_size=batch_size, net=net, transformer=transformer))
mean_embedding = np.mean(embeddings, axis=0)


    

    




CPU times: user 2min 46s, sys: 16.2 s, total: 3min 2s
Wall time: 1min 29s

In [19]:

    

orig_embedding = next(get_features(
    [caffe.io.load_image(img)],
    layers=layers, batch_size=batch_size, net=net, transformer=transformer))


    

In [20]:

    

import requests


    

In [21]:

    

with open('all_images') as f:
    file_names = [s.strip() for s in f]


    

In [ ]:

    

orig_neighbors = requests.post('http://low1:5000/query/10',
      json={'features': [float(f) for f in orig_embedding]})


    

In [ ]:

    

mean_neighbors = requests.post('http://low1:5000/query/10',
      json={'features': [float(f) for f in mean_embedding]})


    

In [ ]:

    

for i in orig_neighbors:
    display(file_names[i])
    display(Image(file_names[i]))


    

In [ ]:

    

for i in mean_neighbors:
    display(file_names[i])
    display(Image(file_names[i]))