Extract dense optical flow and save as grayscale or RGB images

Readme.md

#Extract dense optical flow and save as grayscale or RGB images Copyright @ Kai Kang ([email protected]) 2016

##Dependencies

• OpenCV 2 with Python interface
• glob, numpy, siopy

##Usages

# help messages
$ python gen_optical_flow.py -h

##Known issues

• Generating false optical flows on some Linux distributions, not sure of the bugs. Haven't found any bugs on OSX. Please use it carefully.

gen_optical_flow.py

#!/usr/bin/env python
import argparse
import cv2
import os
import glob
import sys
import numpy as np
import scipy.io as sio
import time

def cvReadGrayImg(img_path):
    return cv2.cvtColor(cv2.imread(img_path), cv2.COLOR_BGR2GRAY)

def saveOptFlowToImage(flow, basename, merge):
    if merge:
        # save x, y flows to r and g channels, since opencv reverses the colors
        cv2.imwrite(basename+'.png', flow[:,:,::-1])
    else:
        cv2.imwrite(basename+'_x.JPEG', flow[...,0])
        cv2.imwrite(basename+'_y.JPEG', flow[...,1])


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('vid_dir')
    parser.add_argument('save_dir')
    parser.add_argument('--bound', type=float, required=False, default=15,
                        help='Optical flow bounding. [-bound, bound] will be mapped to [0, 255].')
    parser.add_argument('--merge', dest='merge', action='store_true',
                        help='Merge optical flow in x and y axes into RGB images rather than saving each to a grayscale image.')
    parser.add_argument('--debug', dest='visual_debug', action='store_true',
                        help='Visual debugging.')
    parser.set_defaults(merge=False, visual_debug=False)
    args = parser.parse_args()

    norm_width = 500.
    bound = args.bound

    images = glob.glob(os.path.join(args.vid_dir,'*'))
    print ("Processing {}: {} files... ".format(args.vid_dir, len(images))),
    sys.stdout.flush()
    tic = time.time()
    img2 = cvReadGrayImg(images[0])
    for ind, img_path in enumerate(images[:-1]):
        img1 = img2
        img2 = cvReadGrayImg(images[ind+1])
        h, w = img1.shape
        fxy = norm_width / w
        # normalize image size
        flow = cv2.calcOpticalFlowFarneback(
            cv2.resize(img1, None, fx=fxy, fy=fxy),
            cv2.resize(img2, None, fx=fxy, fy=fxy),
            0.5, 3, 15, 3, 7, 1.5, 0)
        # map optical flow back
        flow = flow / fxy
        # normalization
        flow = np.round((flow + bound) / (2. * bound) * 255.)
        flow[flow < 0] = 0
        flow[flow > 255] = 255
        flow = cv2.resize(flow, (w, h))

        # Fill third channel with zeros
        flow = np.concatenate((flow, np.zeros((h,w,1))), axis=2)

        # save
        if not os.path.isdir(args.save_dir):
            os.makedirs(args.save_dir)
        basename = os.path.splitext(os.path.basename(img_path))[0]
        saveOptFlowToImage(flow, os.path.join(args.save_dir, basename), args.merge)

        if args.visual_debug:
            mag, ang = cv2.cartToPolar(flow[...,0], flow[...,1])
            hsv = np.zeros_like(cv2.imread(img_path))
            hsv[...,1] = 255
            hsv[...,0] = ang*180/np.pi/2
            hsv[...,2] = cv2.normalize(mag,None,0,255,cv2.NORM_MINMAX)
            bgr = cv2.cvtColor(hsv,cv2.COLOR_HSV2BGR)

            cv2.imshow('optical flow',bgr)
            k = cv2.waitKey(30) & 0xff
            if k == 27:
                break

    # duplicate last frame
    basename = os.path.splitext(os.path.basename(images[-1]))[0]
    saveOptFlowToImage(flow, os.path.join(args.save_dir, basename), args.merge)
    toc = time.time()
    print "{:.2f} min, {:.2f} fps".format((toc-tic) / 60., 1. * len(images) / (toc - tic))