AI Paper Atlas

Research map / Video / Action recognition

DeFB: Decomposed Feature Learning for Real-Time Multi-Person Eyeblink Detection in Untrimmed In-the-Wild Videos

Jinfang Gan, Wenzheng Zeng, Yang Xiao, Xintao Zhang, Chaoyang Zheng, Ran Zhao, Ran Wang, Min Du, Zhiguo Cao

AAAI 2026 · Video › Action recognition · approach: Fully supervised temporal action detection · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
add decomposed feature space (coarse face features plus dedicated fine-grained eye module) on top of async training (ablation)DeFB · MPEblinkBlink-AP0.5:0.95: 12.93 -> 24.65Improved
add asynchronous training on top of decomposed features (ablation)DeFB · MPEblinkBlink-AP0.5:0.95: 14.47 -> 24.65Improved
add both decomposed features and asynchronous training (ablation)InstBlink-style unified model · MPEblinkBlink-AP0.5:0.95: 7.96 -> 24.65Improved
add dynamic proposal selection instead of rasterized uniform proposals (ablation)DeFB · MPEblinkInst-AP: 74.63 -> 76.07Improved
drop the high-resolution encoder scale for face modeling (toy experiment) (ablation)InstBlink · MPEblinkInst-AP: -1.34%No clear effect
add separate lightweight spatio-temporal module for eyeblink (toy experiment) (ablation)InstBlink · MPEblinkBlink-AP: +39.47%Improved
decompose feature space, dropping the fine-grained scale for faces (toy experiment) (ablation)InstBlink · MPEblinkefficiency: +14.28%Improved
plug DeFB into general action detectors (ActionFormer, TriDet, DyFADet, DiGIT)action detectors · MPEblinkBlink-AP: 1.42 -> 26.36 (Actionformer)Improved
DeFB instead of InstBlink for cross-dataset transferunified eyeblink detectors · HUST-LEBW (trained on MPEblink)F1: 83.45 -> 86.92Improved
DeFB instead of InstBlink on HUST-LEBW recallunified eyeblink detectors · HUST-LEBW (trained on MPEblink)Recall: 91.34 -> 88.98Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 3 of 3 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy