actionformer_release

Why we included this project

ActionFormer is one of the first Transformer-based models for temporal action localization, and this repo is the reference implementation from the ECCV 2022 paper. The model finds when an action starts and ends in untrimmed video and names the action, all in a single pass, with no proposal stage and no pre-defined anchor windows. The repo includes training and evaluation code for THUMOS14, ActivityNet 1.3, and EPIC-Kitchens 100, with configs, data loaders, and pre-trained checkpoints, so the reported results are reproducible without reverse-engineering the setup. It also served as the backbone for many winning entries in the Ego4D Moment Queries Challenge, which makes it a practical base for moment localization and video grounding work. The code is clean and Detectron2-inspired, so researchers comparing temporal action localization methods, or teams building video understanding systems, should find it easy to adapt.