app.py

import os
os.system("git clone https://github.com/google-research/frame-interpolation")
import sys
sys.path.append("frame-interpolation")

import math
import cv2
import numpy as np
import tensorflow as tf
import mediapy
from PIL import Image

import gradio as gr

from huggingface_hub import snapshot_download
from image_tools.sizes import resize_and_crop


model = snapshot_download(repo_id="akhaliq/frame-interpolation-film-style")
from eval import interpolator, util
interpolator = interpolator.Interpolator(model, None)

ffmpeg_path = util.get_ffmpeg_path()
mediapy.set_ffmpeg(ffmpeg_path)


def do_interpolation(frame1, frame2, interpolation, n):
    print("tween frames: " + str(interpolation))
    print(frame1, frame2)
    input_frames = [frame1, frame2]
    frames = list(
        util.interpolate_recursively_from_files(
            input_frames, int(interpolation), interpolator))
    
    #print(frames)
    mediapy.write_video(f"{n}_to_{n+1}_out.mp4", frames, fps=25)
    return f"{n}_to_{n+1}_out.mp4"
    
def get_frames(video_in, step, name, n):
    frames = []
    cap = cv2.VideoCapture(video_in)
    cframes = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    cfps = int(cap.get(cv2.CAP_PROP_FPS))
    print(f'frames: {cframes}, fps: {cfps}')

    #resize the video
    #clip = VideoFileClip(video_in)
    
    #check fps
    #if cfps > 25:
    #    print("video rate is over 25, resetting to 25")
    #    clip_resized = clip.resize(height=1024)
    #    clip_resized.write_videofile("video_resized.mp4", fps=25)
    #else:
    #    print("video rate is OK")
    #    clip_resized = clip.resize(height=1024)
    #    clip_resized.write_videofile("video_resized.mp4", fps=cfps)
    
    #print("video resized to 1024 height")
    
    # Opens the Video file with CV2
    #cap = cv2.VideoCapture("video_resized.mp4")
    
    fps = cap.get(cv2.CAP_PROP_FPS)
    print("video fps: " + str(fps))
    i=0
    while(cap.isOpened()):
        ret, frame = cap.read()
        if ret == False:
            break
        #if resize_w > 0:
          #resize_h = resize_w / 2.0
          #frame = cv2.resize(frame, (int(resize_w), int(resize_h)))
        
        cv2.imwrite(f"{str(n)}_{name}_{step}{str(i)}.png", frame)
        frames.append(f"{str(n)}_{name}_{step}{str(i)}.png")
        i+=1
    
    cap.release()
    cv2.destroyAllWindows()
    print("broke the video into frames")
    
    return frames, fps


def create_video(frames, fps, type):
    print("building video result")
    imgs = []
    for j, img in enumerate(frames):
        imgs.append(cv2.cvtColor(cv2.imread(img).astype(np.uint8), cv2.COLOR_BGR2RGB))

    mediapy.write_video(type + "_result.mp4", imgs, fps=fps)
    return type + "_result.mp4"

    
def infer(f_in, interpolation, fps_output):
    
    fps_output = logscale(fps_output)
    # 1. break video into frames and get FPS
    #break_vid = get_frames(url_in, "vid_input_frame", "origin", resize_n)
    frames_list = f_in #break_vid[0]
    fps = 1 #break_vid[1]
    print(f"ORIGIN FPS: {fps}")
    n_frame = int(300) #limited to 300 frames
    #n_frame = len(frames_list)
    
    if n_frame >= len(frames_list):
        print("video is shorter than the cut value")
        n_frame = len(frames_list)
    
    # 2. prepare frames result arrays
    result_frames = []
    print("set stop frames to: " + str(n_frame))

    for idx, frame in enumerate(frames_list[0:int(n_frame)]):
        if idx < len(frames_list) - 1:
            next_frame = frames_list[idx+1]

            interpolated_frames = do_interpolation(frame, next_frame, interpolation, idx) # should return a list of interpolated frames
            break_interpolated_video = get_frames(interpolated_frames, "interpol", f"{idx}_", -1)
            print(break_interpolated_video[0])
            for j, img in enumerate(break_interpolated_video[0][0:len(break_interpolated_video[0])-1]):
                print(f"IMG:{img}")
                os.rename(img, f"{idx}_to_{idx+1}_{j}.png")
                result_frames.append(f"{idx}_to_{idx+1}_{j}.png")
            
            print("frames " + str(idx) + " & " + str(idx+1) + "/" + str(n_frame) + ": done;")
            #print(f"CURRENT FRAMES: {result_frames}")
    result_frames.append(f"{frames_list[n_frame-1]}")
    final_vid = create_video(result_frames, fps_output, "interpolated")

    files = final_vid
    print("interpolated frames: " + str(len(frames_list)) + " -> " + str(len(result_frames)))
    cv2.destroyAllWindows()

    return final_vid, files


def remove_bg(fl, s, l, v):
    frame = cv2.imread(fl).astype(np.uint8)
    
    b = 5
    #subtract background (get scene with shadow)
    bg = cv2.medianBlur(frame, 255)
    
    frame_ = ((bg.astype(np.int16)-frame.astype(np.int16))+127).astype(np.uint8)
    frame_ = cv2.bilateralFilter(frame_, b*4, b*8, b*2)
    frame_ = cv2.medianBlur(frame_, b)
    
    element = cv2.getStructuringElement(cv2.MORPH_RECT, (2*b+1, 2*b+1), (b,b))
    frame_ = cv2.erode(cv2.dilate(frame_, element), element)

    
    #correct hue against light
    bg_gray = cv2.cvtColor(cv2.cvtColor(bg, cv2.COLOR_BGR2GRAY), cv2.COLOR_GRAY2BGR)
    bg_diff = (bg-bg_gray).astype(np.int16)
    frame_c = (frame.astype(np.int16)-bg_diff).astype(np.uint8)


    edges = cv2.Laplacian(cv2.cvtColor(frame, cv2.COLOR_BGR2GRAY), cv2.CV_64F)
    blur_s = np.zeros_like(edges)
    for i in range(2, frame.shape[0]-2):
        for j in range(2, frame.shape[1]-2):
            d = edges[i-2:i+2,j-2:j+2].var()
            blur_s[i,j] = (d/512).astype(np.uint8)

    #remove regions of low saturation and lightness (get scene without shadow)
    m = cv2.inRange(cv2.cvtColor(frame_c, cv2.COLOR_RGB2HSV), np.array([0,0,0]), np.array([180,s,l]))
    print(v)
    mask = cv2.inRange(blur_s, 0, v)
    masks = np.bitwise_and(m, mask)
    frame[masks>0] = (127,127,127)
    frame = cv2.medianBlur(frame, b)
    
    
    m_ = frame_.reshape((-1,3))
    # convert to np.float32
    m_ = np.float32(m_)
    # define criteria, number of clusters(K) and apply kmeans()
    criteria = (cv2.TERM_CRITERIA_EPS + cv2.TERM_CRITERIA_MAX_ITER, 16, 1.0)
    K = 3
    ret,label,center=cv2.kmeans(m_, K, None, criteria, 16, cv2.KMEANS_PP_CENTERS)
    # Now convert back into uint8, and make original image
    center = np.uint8(center)
    res = center[label.flatten()]
    frame_ = res.reshape((frame_.shape))
    
    
    #remove shadows at edges
    m_ = cv2.inRange(frame_, np.array([128,128,128]), np.array([255,255,255]))
    frame_[m_>0] = (255,255,255)
    cv2.rectangle(frame_,(0,0),(frame_.shape[1]-1,frame_.shape[0]-1),(255,255,255),7)
    mask = cv2.floodFill(frame_, None, (0, 0), 255, 0, 0, (4 | cv2.FLOODFILL_FIXED_RANGE))[2] #(4 | cv2.FLOODFILL_FIXED_RANGE | cv2.FLOODFILL_MASK_ONLY | 255 << 8)
    # 255 << 8 tells to fill with the value 255)
    mask = mask[1:mask.shape[0]-1, 1:mask.shape[1]-1]
    frame_[mask>0] = (127,127,127)
    m_ = cv2.inRange(frame_, np.array([0,0,0]), np.array([127,127,127]))
    frame_[m_>0] = (127,127,127)
    

    cv2.imwrite(fl, frame) #frame_
    return fl

def logscale(linear):
    return int(math.pow(2, linear))

def linscale(linear):
    return int(math.log2(linear))

def sharpest(fl, i):
    break_vid = get_frames(fl, "vid_input_frame", "origin", i)

    frames = []
    blur_s = []
    for jdx, fr in enumerate(break_vid[0]):
        frames.append(cv2.imread(fr).astype(np.uint8))
        blur_s.append(cv2.Laplacian(cv2.cvtColor(frames[len(frames)-1], cv2.COLOR_BGR2GRAY), cv2.CV_64F).var())
        print(str(int(blur_s[jdx])))

    indx = np.argmax(blur_s)
    fl = break_vid[0][indx]

    n = 25
    half = int(n/2)
    if indx-half < 0:
        n = indx*2+1
    elif indx+half >= len(frames):
        n = (len(frames)-1-indx)*2+1
        
    #denoise
    frame = cv2.fastNlMeansDenoisingColoredMulti(
        srcImgs = frames,
        imgToDenoiseIndex = indx,
        temporalWindowSize = n,
        hColor = 5,
        templateWindowSize = 21,
        searchWindowSize = 21)

    cv2.imwrite(fl, frame)
    print(str(i) +'th file, sharpest frame: '+str(indx)+', name: '+fl)
    return fl

def sortFiles(e):
    e = e.split('/')
    return e[len(e)-1]

def loadf(f, s, l, v):
    if f != None and f[0] != None:
        f.sort(key=sortFiles)
        fnew = []

        for i, fl in enumerate(f):
            ftype = fl.split('/')
            if ftype[len(ftype)-1].split('.')[1] == 'mp4':
                fl = sharpest(fl, i)

            fl = remove_bg(fl, s, l, v)
            fnew.append(fl)

        return fnew, fnew
    else:
        return f, f


title="""
<div style="text-align: center; max-width: 500px; margin: 0 auto;">
        <div
        style="
            display: inline-flex;
            align-items: center;
            gap: 0.8rem;
            font-size: 1.75rem;
            margin-bottom: 10px;
        "
        >
        <h1 style="font-weight: 600; margin-bottom: 7px;">
            Video interpolation from images with FILM
        </h1>
        
        </div>
       <p> This space uses FILM to generate interpolation frames in a set of image files you need to turn into a video.
       Limited to 300 uploaded frames, from the beginning of your input.<br />
       <a style="display:inline-block" href="https://huggingface.co/spaces/freealise/video_frame_interpolation?duplicate=true"><img src="https://img.shields.io/badge/-Duplicate%20Space-blue?labelColor=white&style=flat&logo=data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAABAAAAAQCAYAAAAf8/9hAAAAAXNSR0IArs4c6QAAAP5JREFUOE+lk7FqAkEURY+ltunEgFXS2sZGIbXfEPdLlnxJyDdYB62sbbUKpLbVNhyYFzbrrA74YJlh9r079973psed0cvUD4A+4HoCjsA85X0Dfn/RBLBgBDxnQPfAEJgBY+A9gALA4tcbamSzS4xq4FOQAJgCDwV2CPKV8tZAJcAjMMkUe1vX+U+SMhfAJEHasQIWmXNN3abzDwHUrgcRGmYcgKe0bxrblHEB4E/pndMazNpSZGcsZdBlYJcEL9Afo75molJyM2FxmPgmgPqlWNLGfwZGG6UiyEvLzHYDmoPkDDiNm9JR9uboiONcBXrpY1qmgs21x1QwyZcpvxt9NS09PlsPAAAAAElFTkSuQmCC&logoWidth=14" alt="Duplicate Space"></a> 
       </p>
    </div>
"""

with gr.Blocks() as demo:
    with gr.Column():
        gr.HTML(title)
        with gr.Row():
            with gr.Column():
                with gr.Accordion(label="Upload files here", open=True):
                    files_input = gr.File(file_count="multiple", file_types=['image', '.mp4'])
                gallery_input = gr.Gallery(label="Slideshow", preview=True, columns=8192, interactive=False)
                with gr.Accordion(label="Background removal settings", open=False):
                    max_ = gr.Label(value="Shadow maximums:")
                    max_s = gr.Slider(minimum=0, maximum=255, step=1, value=32, label="Saturation")
                    max_l = gr.Slider(minimum=0, maximum=255, step=1, value=64, label="Lightness")
                    max_v = gr.Slider(minimum=0, maximum=255, step=1, value=16, label="Variance")
                files_input.upload(fn=loadf, inputs=[files_input, max_s, max_l, max_v], outputs=[files_input, gallery_input])
                with gr.Row():
                    interpolation_slider = gr.Slider(minimum=1, maximum=5, step=1, value=1, label="Interpolation Steps: ")
                    interpolation = gr.Number(value=1, show_label=False, interactive=False)
                    interpolation_slider.change(fn=logscale, inputs=[interpolation_slider], outputs=[interpolation])
                with gr.Row():
                    fps_output_slider = gr.Slider(minimum=0, maximum=5, step=1, value=0, label="FPS output: ")
                    fps_output = gr.Number(value=1, show_label=False, interactive=False)
                    fps_output_slider.change(fn=logscale, inputs=[fps_output_slider], outputs=[fps_output])
                submit_btn = gr.Button("Submit")
            
            with gr.Column():
                video_output = gr.Video()
                file_output = gr.File()
    
    gr.Examples(
        examples=[[
            ["./examples/0.png", "./examples/1.png", "./examples/2.png", "./examples/3.png", "./examples/4.png"], 
            32, 64, 16
        ]],
        fn=loadf,
        inputs=[files_input, max_s, max_l, max_v],
        outputs=[files_input, gallery_input],
        cache_examples=True
    )
    
    submit_btn.click(fn=infer, inputs=[files_input, interpolation_slider, fps_output_slider], outputs=[video_output, file_output])

demo.launch()