EffectErase: Joint Video Object Removal and Insertion with a Reciprocal Learning Scheme

Forum topic · 小凯 · 2026-03-22

Summary

Researchers Yang Fu, Yike Zheng, and Ziyun Dai introduce VOR, a large-scale dataset of 60K high-quality video pairs designed for video object removal, alongside EffectErase, an effect-aware video object removal method. The key insight of EffectErase is to treat video object insertion as the inverse auxiliary task within a reciprocal learning scheme, allowing the two tasks to reinforce each other and improve removal quality. The work, posted on arXiv as 2503.16887 and categorized under computer vision, addresses the scarcity of high-quality paired training data for video object removal by providing a curated dataset at scale. By jointly modeling removal and insertion, the method aims to produce high-quality results where removed objects leave no visual traces, such as shadows or reflections. This post is a brief bilingual summary originally shared on zhichai.net on March 19, 2026, and automatically collected on March 22, 2026.

Paper Overview

Field: Computer Vision (CV) Authors: Yang Fu, Yike Zheng, Ziyun Dai Published: 2026-03-19 arXiv: 2503.16887

Abstract (English)

We introduce VOR, a large-scale dataset with 60K high-quality video pairs for video object removal. Building on VOR, we propose EffectErase, an effect-aware video object removal method that treats video object insertion as the inverse auxiliary task within a reciprocal learning scheme.

Key Contributions

VOR dataset: a large-scale dataset containing 60K high-quality video pairs specifically curated for the video object removal task.
EffectErase: an effect-aware video object removal method.
Reciprocal learning: video object insertion is treated as the inverse auxiliary task of removal, and the two tasks are learned jointly in a reciprocal scheme to boost removal quality.

Notes

The Chinese abstract in the source post is a direct translation of the English abstract above.
Auto-collected from zhichai.net on 2026-03-22.