XMem
Long-video object segmentation with a three-tier memory modeled on human Atkinson-Shiffrin memory.
About
XMem solves the problem that breaks most video object segmentation models: long videos. It structures memory like the Atkinson-Shiffrin model of human memory, with fast sensory memory, a working memory of recent frames, and a compacted long-term memory, so GPU usage stays bounded no matter how long the video runs. Give it a mask on the first frame and it tracks and segments the object through occlusions and appearance changes.
From UIUC (Ho Kei Cheng), published at ECCV 2022 under the MIT license, XMem became the standard long-term VOS baseline and the backbone for follow-ups like Cutie and DEVA. It runs around 20 FPS on long videos and ships with an interactive GUI demo. Its successor Cutie beats it on accuracy, but XMem remains the reference design for memory-efficient video segmentation.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Author-reported (ECCV 2022): strong DAVIS-2017 val/test-dev J&F; ~20 FPS on long videos with bounded GPU memory
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: ECCV 2022 paper (arXiv:2207.07115); MIT license and repo activity confirmed via the official GitHub repo metadata. Strengths: bounded-memory long-video design, clean codebase, GUI demo, spawned Cutie/DEVA. Docked for: (1) accuracy superseded by Cutie on standard benchmarks; (2) single-object-per-pass workflow is slower for multi-object jobs. License: MIT.
How the trust score is built
Full methodology →- Adoption · stars/downloads, log-scaled vs peers
- 7.0/10
- Maintenance · recency, releases, issue load
- 5.0/10
- Openness · license + docs
- 9.0/10
- Manual review · sanity read + red-flag check
- 7.0/10
- · ECCV 2022 (arXiv 2207.07115, Jul 2022)
- · ~20 FPS on long videos with bounded GPU memory, author-reported
- · Standard long-term VOS baseline; backbone for Cutie and DEVA