Add flow_to_image() visualization util

[NicolasHug]

Now that RAFT is done, it'd be nice to have a conversion util from a 2xHxW flow to a 3xHxW RGB image. I think the implementation of the original RAFT paper follows a standard color map, so we could do something similar (trimmed down to the minimal parts, e.g. no BRG etc.) https://github.com/princeton-vl/RAFT/blob/master/core/utils/flow_viz.py

Note: this is a visualization util meant to land in torchvision.utils; it's not meant to be a transform nor an ops.

[oke-aditya]

I can take this 😄

[NicolasHug]

Thanks @oke-aditya ! Looking forward to your PR!

[oke-aditya]

It would take some time for me. As I haven't tried raft. Not experienced with optical flow either.

But definitely will prototype before end of year.

[NicolasHug]

I'll be on holidays the last 2 weeks of Dec, so there's absolutely no rush on our side ;)

Just FYI, you don't really need to be familiar with optical flow models or with RAFT for this task, since it's just about converting a flow into an RGB image.

But if you want to play around with RAFT, I've been using this (hacky) script to apply our model to videos:

import os
from argparse import ArgumentParser
import sys
sys.path.append('../RAFT/core')  # This is the original RAFT repo from https://github.com/princeton-vl/RAFT

import cv2
import numpy as np
import torch

from torchvision.prototype.models.optical_flow import raft_large
from utils import flow_viz


def frame_preprocess(frame, device):
    frame = torch.from_numpy(frame).permute(2, 0, 1).float()
    frame = frame.unsqueeze(0)
    frame = frame.to(device)
    return frame


def vizualize_flow(img, flo, save, counter):
    # permute the channels and change device is necessary
    img = img[0].permute(1, 2, 0).cpu().numpy()
    flo = flo[0].permute(1, 2, 0).cpu().numpy()

    # map flow to rgb image
    flo = flow_viz.flow_to_image(flo)
    flo = cv2.cvtColor(flo, cv2.COLOR_RGB2BGR)

    # concatenate, save and show images
    img_flo = np.concatenate([img, flo], axis=0)
    if save:
        cv2.imwrite(f"demo_frames/frame_{str(counter)}.jpg", img_flo)
    cv2.imshow("Optical Flow", img_flo / 255.0)
    k = cv2.waitKey(25) & 0xFF
    if k == 27:
        return False
    return True


def inference(args):
    model = raft_large(weights="Raft_Large_Weights.C_T_SKHT_K_V2")

    save = args.save
    if save:
        if not os.path.exists("demo_frames"):
            os.mkdir("demo_frames")

    device = "cpu"
    model.eval()

    video_path = args.video

    cap = cv2.VideoCapture(video_path)
    cap.set(cv2.CAP_PROP_POS_FRAMES, int(3 * 30))
    ret, frame_1 = cap.read()

    frame_1 = frame_preprocess(frame_1, device)

    counter = 0
    with torch.no_grad():
        while True:
            # read the next frame
            ret, frame_2 = cap.read()
            if not ret:
                break
            frame_2 = frame_preprocess(frame_2, device)
            flow_preds = model(2 * frame_1 / 255 - 1, 2 * frame_2 / 255 - 1)
            flow_up = flow_preds[-1]
            ret = vizualize_flow(frame_1, flow_up, save, counter)
            if not ret:
                break
            frame_1 = frame_2
            counter += 1


def main():
    parser = ArgumentParser()
    parser.add_argument("--model", help="restore checkpoint")
    parser.add_argument("--iters", type=int, default=12)
    parser.add_argument("--video", type=str, default="./videos/car.mp4")
    parser.add_argument("--save", action="store_true", help="save demo frames")
    parser.add_argument("--small", action="store_true", help="use small model")
    parser.add_argument(
        "--mixed_precision", action="store_true", help="use mixed precision"
    )

    args = parser.parse_args()
    inference(args)


if __name__ == "__main__":
    main()

(adapted from https://github.com/spmallick/learnopencv/blob/master/Optical-Flow-Estimation-using-Deep-Learning-RAFT/inference.py). The goal of this issue is to implement the flow_viz.flow_to_image() of the script

[oke-aditya]

I'll be on holidays the last 2 weeks of Dec, so there's absolutely no rush on our side ;)

Happy Holidays Nicolas. :)

I will have a look at the script and try to implement a minimalist version.
I will ping you back in a PR hopefully soon.