Video-Stitching-Perspective-Transition

Homography-based perspective video transition using LoFTR feature matching and DeepLabV3 person segmentation for seamless foreground-aware stitching between clips.

Overview

This project generates a perspective-aligned transition between two videos so that a hard cut appears as a continuous camera movement.

It works by:

• Matching features between the last frame of Video A and a reference frame of Video B using LoFTR
• Estimating a homography with RANSAC
• Warping Video B into Video A’s perspective
• Segmenting people using DeepLabV3
• Blending foreground and background with temporal fade control
• Rendering a final stitched output

Requirments

• Python 3.10+

torch>=2.0
torchvision>=0.15
opencv-python>=4.8
kornia>=0.7
numpy>=1.23

Install with:

pip install -r requirements.txt

CUDA Support

Install PyTorch with CUDA from:

pip install torch torchvision --index-url https://download.pytorch.org/whl/cu121

Running the script

python video_stitcher.py -a path/to/videoA.mp4 -b path/to/videoB.mp4 -o output.mp4 --overlap 42 --loftr-max-dim 1152 --fade-in 10 --fade-out 10

OR simply go with default command:

python video_stitcher.py -a path/to/videoA.mp4 -b path/to/videoB.mp4 -o output.mp4

Arguments

Argument	Short	Type	Default	Description
--video-a	-a	str	pre.mp4	Path to first video clip (Outgoing)
--video-b	-b	str	post.mp4	Path to second video clip (Incoming)
--output	-o	str	transition.mp4	Output video path
--overlap		int	40	Number of frames to overlap/transition between videos
--loftr-max-dim		int	1152	Maximum dimension for LoFTR feature matching
--fade-in		int	10	Number of frames to fade in the foreground at the start
--fade-out		int	10	Number of frames to fade out the pre-clip at the end

Use the video_stitcher_layers.py to get separate outputs synced layers for compositing.

Fix download issues:

urllib.error.URLError: <urlopen error [SSL: CERTIFICATE_VERIFY_FAILED] certificate verify failed: unable to get local issuer certificate (_ssl.c:1000)>

If you encounter this error while downloading loftr_indoor_ds_new.ckpt, simply download it manually instead (this happens due to some HTTP issues).

Steps

• Download the indoor_ds_new.ckpt from this drive link: Google Drive
• Rename the file from indoor_ds_new.ckpt to loftr_indoor_ds_new.ckpt
• Paste the file in C:\Users\user\.cache\torch\hub\checkpoints

That's it, it will solve the error. (This is not an issue with other models)

Example

PRE CLIP	POST CLIP	STITCHED CLIP

Tip: You can change the Segmentation model to generate better masks. Check Here

License: MIT

Copyright (c) 2026 Akash Bora
Get more video effects at www.akascape.com 👈