Example #1
0
def create_gist_database(data_dir, img_dirs):
    """Create a GIST database"""

    if (not os.path.isfile(gists_db_name)) or (not os.path.isfile(f_db_name)):

        gist_data = []
        file_names = []
        for dir in img_dirs:
            scene_dir = os.path.join(data_dir, dir)
            img_files = [f for f in os.listdir(scene_dir)
                         if os.path.isfile(os.path.join(scene_dir, f))
                         and 'jpg' in f]

            for f in img_files:
                img_file = os.path.join(scene_dir, f)

                print img_file
                gist, param \
                    = lmgist.lmgist(img_file, param)

                gist_data.append(gist)
                file_names.append(img_file)

        np.save(gists_db_name, gist_data)
        np.save(f_db_name, file_names)

        with open('gistparams', 'w') as f:
            pickle.dump(param, f)
    else:
        print '%s and %s already exist' % (gists_db_name, f_db_name)

    return
Example #2
0
def get_gist(query_name, img_query, img_mask, param):
    '''
    Get the gist descriptor for the query image with mask
    '''

    # get the gist parameter of the source file
    gist, param = lmgist.lmgist(query_name, param)

    # resize the query and mask images to match the one used to compute
    # the GIST descriptor
    img_query = transform.resize(img_query, (param.img_size, param.img_size),
                                 clip=True, order=1)
    img_mask = transform.resize(img_mask, (param.img_size, param.img_size),
                                clip=True, order=1)

    # make sure the mask is 2d and in [0, 1]
    if len(img_mask.shape) != 2:
        raise TypeError('Mask image needs to be 2D!')
    elif not ((np.min(img_mask) == 0) and (np.max(img_mask == 1))):
        img_mask = (img_mask - np.min(img_mask)) \
            / (np.max(img_mask) - np.min(img_mask))

    # GIST weights are the average value of mask pixels within each block
    s = (np.linspace(0, param.img_size, param.number_blocks + 1)).astype(int)
    block_weight = np.zeros((param.number_blocks, param.number_blocks))
    for y in range(param.number_blocks):
        for x in range(param.number_blocks):
            block = img_mask[s[y]:s[y + 1], s[x]:s[x + 1]]
            block_weight[y, x] = np.mean(block)

    block_weight = 1 - block_weight
    n_filters = sum(param.orientations_per_scale)
    block_weight = matlib.repmat(
        np.reshape(block_weight.T, [param.number_blocks ** 2, 1]), n_filters, 1)

    return gist, block_weight