Goal of this post is to investigate Action Detection datasets such as AVA Action and AVA-Kinetics, as well as the Homage Dataset.
내용
- Explore ActivityNet Challenges
- Conducted initial research to grasp the trends by examining various datasets and tasks related to video understanding.
- findings: I Decided to look into AVA-Kinetics and Homage.
By the way, since 2019 that SlowFast won ActivityNet Challenges, Chinese groups have been consistently securing the 1st place in 20, 21, and 22.
- This year (2022) focuses on active speaker detection, which is different due to the use of audio.
- 1st place: Project Page, Code to be released.
- 2nd place by Intel: Paper
- About AVA-Kinetics
- Spatio-Temporal Action Localization Task
- Research on state-of-the-art (SOTA): paperswithcode, List of AVA Challenge Winners
- "RM" outperformed the ACAR-Net (CVPR 2021 paper) which is a winning model of AVA Action (ActivityNet Challenge 2020)
- RM: This model won the AVA-Kinetics 2021 Challenge. paper, but the GitHub code is not available.
- About Homage
About Model
- It seems meaningless to search for SOTA on paperswithcode as there are no notable models reported.
- Official GitHub page provides detailed information about the dataset github
- Let's check the challenge results. 2021 Results
- The 2022 results seem unavailable as the challenge ended in June. Challenge Page on CodaLab
About Data
- Dataset Paper link
