#ECCV 2022

The DIS project highlights significant innovations in dichotomous image segmentation, including the introduction of the DIS5K V1.0 dataset and upcoming DIS V2.0 dataset with broader category inclusion. Recognized at ECCV 2022, it features an optimized IS-Net model suited for general use, available through platforms like Huggingface and Bohrium for live demonstrations. The project supports diverse applications from 3D modeling to AR, driven by advanced IS-Net architecture and human correction efforts, demonstrating its role in advancing image processing capabilities.

ActionFormer utilizes a Transformer-based model for effective temporal action localization, achieving 71.0% mAP on THUMOS14. It incorporates local self-attention for temporal context modeling in videos and has demonstrated strong performance across various benchmarks such as ActivityNet and EPIC-Kitchens 100. The model's capabilities are further validated in the Ego4D Moment Queries Challenge. Users can explore ActionFormer's open-source implementation to replicate these outcomes, utilizing the available pre-trained models and configurations suited for major video datasets.