Draw a box, delete anything: vintage video inpainting

What it does Feed it a video and draw a bounding box around the unwanted object in the first frame. The pipeline uses SiamMask to track and segment that object across every subsequent frame, then feeds those masks into a Deep Video Inpainting model to reconstruct the background. You get back a video where the object has been painted out.

The interesting bit This is integration work, not invention: it duct-tapes two distinct CVPR papers—one for visual tracking, one for video inpainting—into a single offline workflow. The quiet value is in the mask-dilation knob; by expanding the segmentation mask slightly, it keeps the inpainter from smearing artifacts at the object’s edges.

Key highlights

• Interactive removal from a single bounding-box initialization.
• Supports both video files and image sequences as input.
• Tunable --mask-dilation parameter to mitigate edge artifacts.
• Runs fully offline using pretrained models.
• Explicitly built atop SiamMask and Deep-Video-Inpainting.

Caveats

• Dependency stack is frozen in time: PyTorch 0.4.0, CUDA 8.0, Python 3.5, and Ubuntu 16.04 are the tested environment, so expect archaeology to run it today.
• The repository is a demo and integration layer rather than a novel model or training framework.

Verdict Worth a look if you need a self-contained, offline proof-of-concept for video object removal and can tolerate legacy dependencies. Avoid if you need a maintained, modern codebase or a trainable architecture.