data_augmentation.py

import tensorflow as tf
import numpy as np

def random_crop(img_list, crop_h, crop_w):
    img_size = tf.shape(img_list[0])
    # crop image and flow
    rand_offset_h = tf.random_uniform([], 0, img_size[0]-crop_h+1, dtype=tf.int32)
    rand_offset_w = tf.random_uniform([], 0, img_size[1]-crop_w+1, dtype=tf.int32)
    
    for i, img in enumerate(img_list):
        img_list[i] = tf.image.crop_to_bounding_box(img, rand_offset_h, rand_offset_w, crop_h, crop_w)
    
    return img_list

def flow_vertical_flip(flow):
    flow = tf.image.flip_up_down(flow)
    flow_u, flow_v = tf.unstack(flow, axis=-1)
    flow_v = flow_v * -1
    flow = tf.stack([flow_u, flow_v], axis=-1)
    return flow

def flow_horizontal_flip(flow):
    flow = tf.image.flip_left_right(flow)
    flow_u, flow_v = tf.unstack(flow, axis=-1)
    flow_u = flow_u * -1
    flow = tf.stack([flow_u, flow_v], axis=-1)
    return flow

def random_flip(img_list):
    is_flip = tf.random_uniform([2], minval=0, maxval=2, dtype=tf.int32)
    
    for i in range(len(img_list)):
        img_list[i] = tf.where(is_flip[0] > 0, tf.image.flip_left_right(img_list[i]), img_list[i])
        img_list[i] = tf.where(is_flip[1] > 0, tf.image.flip_up_down(img_list[i]), img_list[i])  
    return img_list

def random_flip_with_flow(img_list, flow_list):
    is_flip = tf.random_uniform([2], minval=0, maxval=2, dtype=tf.int32)
    for i in range(len(img_list)):
        img_list[i] = tf.where(is_flip[0] > 0, tf.image.flip_left_right(img_list[i]), img_list[i])
        img_list[i] = tf.where(is_flip[1] > 0, tf.image.flip_up_down(img_list[i]), img_list[i]) 
    for i in range(len(flow_list)):
        flow_list[i] = tf.where(is_flip[0] > 0, flow_horizontal_flip(flow_list[i]), flow_list[i])
        flow_list[i] = tf.where(is_flip[1] > 0, flow_vertical_flip(flow_list[i]), flow_list[i])  
    return img_list, flow_list


def random_channel_swap(img_list):
    channel_permutation = tf.constant([[0, 1, 2],
                                       [0, 2, 1],
                                       [1, 0, 2],
                                       [1, 2, 0], 
                                       [2, 0, 1],
                                       [2, 1, 0]])    
    rand_i = tf.random_uniform([], minval=0, maxval=6, dtype=tf.int32)
    perm = channel_permutation[rand_i]
    for i, img in enumerate(img_list):
        channel_1 = img[:, :, perm[0]]
        channel_2 = img[:, :, perm[1]]
        channel_3 = img[:, :, perm[2]]
        img_list[i] = tf.stack([channel_1, channel_2, channel_3], axis=-1)
    return img_list

def flow_resize(flow, out_size, is_scale=True, method=0):
    '''
        method: 0 mean bilinear, 1 means nearest
    '''
    flow_size = tf.to_float(tf.shape(flow)[-3:-1])
    flow = tf.image.resize_images(flow, out_size, method=method, align_corners=True)
    if is_scale:
        scale = tf.to_float(out_size) / flow_size
        scale = tf.stack([scale[1], scale[0]])
        flow = tf.multiply(flow, scale)
    return flow