Create Stunning Video Ghosting Effects with PaddlePaddle’s DeepLabV3p Model

Creation Background

Inspired by iconic martial‑arts novels and the striking visual style of modern action films, the author explores how to reproduce the dramatic "ghost" effect seen in fight scenes using deep learning.

The technique is demonstrated on a classic duel from a martial‑arts series, showing how the effect enhances visual storytelling.

Algorithm Implementation Process

Environment : Baidu AIStudio deep‑learning platform, Python 3.7, PaddlePaddle 1.8.4.

Step 1: Install PaddleHub

!pip install paddlehub==1.6.2 -i https://mirror.baidu.com/pypi/simple

Step 2: Import required packages

import numpy as np
import os
import paddlehub as hub
import cv2
from moviepy.editor import VideoFileClip
from tqdm import tqdm
import copy

Step 3: Set GPU device (single‑GPU environment) os.environ["CUDA_VISIBLE_DEVICES"] = "0" Step 4: Write inference code that uses the pretrained deeplabv3p_xception65 model to create the ghosting effect

def do_seg(module, frame):
    result = module.segmentation(images=[frame], use_gpu=True)
    return result[0]['data']

module = hub.Module(name="deeplabv3p_xception65_humanseg")
originname = "test.mp4"          # original video
resultname = "test_result.mp4"    # output with ghost effect
shadowcount = 9

cap = cv2.VideoCapture(originname)
fps = int(cap.get(cv2.CAP_PROP_FPS))
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
framecount = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))

fourcc = cv2.VideoWriter_fourcc(*'mp4v')
out = cv2.VideoWriter(resultname, fourcc, fps, (width, height))

tmpres = []
for index in tqdm(range(framecount)):
    ret, frame = cap.read()
    if not ret:
        break
    seg_mask = np.around(do_seg(module, frame) / 255)
    seg_mask3 = np.repeat(seg_mask[:, :, np.newaxis], 3, axis=2)
    background = copy.deepcopy(frame)
    stbackground = copy.deepcopy(frame)
    if len(tmpres) > shadowcount:
        tmpres = tmpres[1:]
    # store current mask and blurred background for later blending
    tmpres.append([
        copy.deepcopy(seg_mask3),
        copy.deepcopy(cv2.GaussianBlur(seg_mask3 * background, (9, 9), 0))
    ])
    tmpres.append([
        copy.deepcopy(seg_mask3),
        copy.deepcopy(seg_mask3 * background)
    ])
    thuman = copy.deepcopy(seg_mask3 * background)
    if index > len(tmpres):
        for fi, [t_mask3, t_human] in enumerate(tmpres):
            background = (
                t_human * (fi + 1) / len(tmpres) +
                t_mask3 * (len(tmpres) - 1 - fi) / len(tmpres) * stbackground +
                (1 - t_mask3) * background
            )
    result = background.astype(np.uint8)
    out.write(result)

cap.release()
out.release()

The deeplabv3p_xception65 model performs semantic segmentation to separate human figures from the background. By blending the current frame with several previously processed frames, a translucent “shadow” trail is created, producing the desired ghosting visual effect.

Actual Test Results

The method was applied to a classic duel scene, demonstrating a convincing ghosting overlay that accentuates motion.