AM40: Enhancing action recognition through matting-driven interaction analysis

Open Webpage

Siqi Liang, Wenxuan Liu, Zhe Li, Kui Jiang, Siyuan Yang, Chia-Wen Lin, Xian Zhong

Published: 01 Jan 2026, Last Modified: 04 Nov 2025Pattern Recognit. 2026EveryoneRevisionsBibTeXCC BY-SA 4.0
Abstract: Highlights•We highlight dynamic actor-background interactions to improve recognition.•We introduce AM40 with alpha mattes for 40 action categories.•We develop the ABD module to decouple background and reduce interference.•We design the SFC module to refine and integrate informative features.•MIR leverages matting to capture interaction cues and boost accuracy.