detect_fake_videos.py

"""
Evaluates a folder of video files or a single file with a xception binary
classification network.

Usage:
python detect_from_video.py
    -i <folder with video files or path to video file>
    -m <path to model file>
    -o <path to output folder, will write one or multiple output videos there>

Author: Andreas Rössler
"""
import os
import argparse
from os.path import join
import cv2
import dlib
import torch
import torch.nn as nn
from PIL import Image as pil_image
from tqdm import tqdm

from network.models import model_selection
from dataset.transform import xception_default_data_transforms


def get_boundingbox(face, width, height, scale=1.3, minsize=None):
    """
    Expects a dlib face to generate a quadratic bounding box.
    :param face: dlib face class
    :param width: frame width
    :param height: frame height
    :param scale: bounding box size multiplier to get a bigger face region
    :param minsize: set minimum bounding box size
    :return: x, y, bounding_box_size in opencv form
    """
    x1 = face.left()
    y1 = face.top()
    x2 = face.right()
    y2 = face.bottom()
    size_bb = int(max(x2 - x1, y2 - y1) * scale)
    if minsize:
        if size_bb < minsize:
            size_bb = minsize
    center_x, center_y = (x1 + x2) // 2, (y1 + y2) // 2

    # Check for out of bounds, x-y top left corner
    x1 = max(int(center_x - size_bb // 2), 0)
    y1 = max(int(center_y - size_bb // 2), 0)
    # Check for too big bb size for given x, y
    size_bb = min(width - x1, size_bb)
    size_bb = min(height - y1, size_bb)

    return x1, y1, size_bb


def preprocess_image(image, cuda=False):
    """
    Preprocesses the image such that it can be fed into our network.
    During this process we envoke PIL to cast it into a PIL image.

    :param image: numpy image in opencv form (i.e., BGR and of shape
    :return: pytorch tensor of shape [1, 3, image_size, image_size], not
    necessarily casted to cuda
    """
    # Revert from BGR
    image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
    # Preprocess using the preprocessing function used during training and
    # casting it to PIL image
    preprocess = xception_default_data_transforms['test']
    preprocessed_image = preprocess(pil_image.fromarray(image))
    # Add first dimension as the network expects a batch
    preprocessed_image = preprocessed_image.unsqueeze(0)
    if cuda:
        preprocessed_image = preprocessed_image.cuda()
    return preprocessed_image


def predict_with_model(image, model, post_function=nn.Softmax(dim=1),
                       cuda=False):
    """
    Predicts the label of an input image. Preprocesses the input image and
    casts it to cuda if required

    :param image: numpy image
    :param model: torch model with linear layer at the end
    :param post_function: e.g., softmax
    :param cuda: enables cuda, must be the same parameter as the model
    :return: prediction (1 = fake, 0 = real)
    """
    # Preprocess
    preprocessed_image = preprocess_image(image, cuda)

    # Model prediction
    output = model(preprocessed_image)
    output = post_function(output)

    # Cast to desired
    _, prediction = torch.max(output, 1)    # argmax
    prediction = float(prediction.cpu().numpy())

    return int(prediction), output


def test_full_image_network(video_path, model_path, output_path, fast,
                            start_frame=0, end_frame=None, cuda=False):
    """
    Reads a video and evaluates a subset of frames with the a detection network
    that takes in a full frame. Outputs are only given if a face is present
    and the face is highlighted using dlib.
    :param video_path: path to video file
    :param model_path: path to model file (should expect the full sized image)
    :param output_path: path where the output video is stored
    :param start_frame: first frame to evaluate
    :param end_frame: last frame to evaluate
    :param cuda: enable cuda
    :return:
    """
    print('Starting: {}'.format(video_path))

    # Read and write
    reader = cv2.VideoCapture(video_path)

    video_fn = video_path.split('/')[-1].split('.')[0]+'.avi'
    os.makedirs(output_path, exist_ok=True)
    fourcc = cv2.VideoWriter_fourcc(*'MJPG')
    fps = reader.get(cv2.CAP_PROP_FPS)
    num_frames = int(reader.get(cv2.CAP_PROP_FRAME_COUNT))
    writer = None

    # Face detector
    face_detector = dlib.get_frontal_face_detector()

    # Load model
    model, *_ = model_selection(modelname='xception', num_out_classes=2)
    if model_path is not None:
        model = torch.load(model_path, map_location="cuda" if torch.cuda.is_available() else "cpu")
        print('Model found in {}'.format(model_path))
    else:
        print('No model found, initializing random model.')
    if cuda:
        model = model.cuda()

    # Text variables
    font_face = cv2.FONT_HERSHEY_SIMPLEX
    thickness = 2
    font_scale = 1

    # Fake frames number
    ff = 0
    ffn = 0

    # Frame numbers and length of output video
    frame_num = 0
    assert start_frame < num_frames - 1
    end_frame = end_frame if end_frame else num_frames
    pbar = tqdm(total=end_frame-start_frame)

    while reader.isOpened():
        _, image = reader.read()
        if image is None:
            break
        if fast:
            frame_num += 10
            pbar.update(10)
        else:
            frame_num+= 1
            pbar.update(1)


        if frame_num < start_frame:
            continue

        # Image size
        height, width = image.shape[:2]

        # Init output writer
        if writer is None:
            writer = cv2.VideoWriter(join(output_path, video_fn), fourcc, fps,
                                     (height, width)[::-1])

        # 2. Detect with dlib
        gray = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
        faces = face_detector(gray, 1)
        if len(faces):
            # For now only take biggest face
            face = faces[0]

            # --- Prediction ---------------------------------------------------
            # Face crop with dlib and bounding box scale enlargement
            x, y, size = get_boundingbox(face, width, height)
            cropped_face = image[y:y+size, x:x+size]

            # Actual prediction using our model
            prediction, output = predict_with_model(cropped_face, model,
                                                    cuda=cuda)

            if prediction == 1:
                ff += 1
            ffn +=1
            # ------------------------------------------------------------------

            # Text and bb
            x = face.left()
            y = face.top()
            w = face.right() - x
            h = face.bottom() - y
            label = 'fake' if prediction == 1 else 'real'
            color = (0, 255, 0) if prediction == 0 else (0, 0, 255)
            output_list = ['{0:.2f}'.format(float(x)) for x in
                           output.detach().cpu().numpy()[0]]
            cv2.putText(image, str(output_list)+'=>'+label, (x, y+h+30),
                        font_face, font_scale,
                        color, thickness, 2)
            # draw box over face
            cv2.rectangle(image, (x, y), (x + w, y + h), color, 2)

        if frame_num >= end_frame:
            break

        writer.write(image)

    pbar.close()
    p = ff / float(ffn) * 100;

    if writer is not None:
        out = {}
        writer.release()
        out["score"] = p
        out["file"] = video_fn
        return out
    else:
        print('Input video file was empty')