|
Google About Google Privacy Terms Feedback |
For announcements and details on upcomming challenges, please sign up to the Google Group: ava-dataset-users.
The 2022 AVA challenge was held as part of the International Challenge on Activity Recognition (ActivityNet) Workshop at CVPR 2022. As in the previous years, the challenge had two separate tasks: AVA-Kinetics for atomic action detection, and Active Speaker for active speaker detection.
This year's Active Speaker winners improved the previous best score by 1.03% absolute mAP - a 15.7% relative error reduction - over an already-strong baseline.
Watch the AVA Challenge 2022 video to see talks from the top-3 teams for each task, and read the details of their approaches in the reports below.
| Rank | Team Name | Entry Report | mAP |
|---|---|---|---|
| 1 | Institute of Computing Technology, Chinese Academy of Sciences & University of Chinese Academy of Sciences | UNICON+: ICTCAS-UCAS Submission to the AVA ActiveSpeaker Task at ActivityNet Challenge 2022 | 94.47 |
| 2 | Intel Labs | Intel Labs at ActivityNet Challenge 2022: SPELL for Long-Term Active Speaker Detection | 93.17 |
| 3 | IVUL-KAUST | IVUL ActivityNet Challenge 2022 Submission - Active Speaker Detection (AVA) | 93.02 |
The 2021 AVA challenge was held as part of the International Challenge on Activity Recognition (ActivityNet) Workshop at CVPR 2021. As in the previous year, the challenge had two separate tasks: AVA-Kinetics for atomic action detection, and Active Speaker for speaker detection.
This year's Active Speaker winners improved the previous best score by 5.6% absolute mAP - an almost 50% relative error reduction - and the AVA-Kinetics winner improved by 1.05% absolute mAP over an already-strong baseline.
Watch the AVA Challenge 2021 video [YouTube] [bilibili] to see talks from the top-3 teams for each task, and read more details about their approaches in the reports below.
| Rank | Team Name | Entry Report | mAP |
|---|---|---|---|
| 1 | Alibaba Group & Tsinghua University | Relation Modeling in Spatio-Temporal Action Localization | 40.67 |
| 2 | Fujitsu | Context Feature for Action Localization | 37.43 |
| 3 | OPPO Research Institute | Pose-Part Network for AVA-Kinetics | 36.90 |
| Rank | Team Name | Entry Report | mAP |
|---|---|---|---|
| 1 | Institute of Computing Technology, Chinese Academy of Sciences & University of Chinese Academy of Sciences & Tomorrow Advancing Life Education Group | ICTCAS-UCAS-TAL Submission to the AVA ActiveSpeaker Task at ActivityNet Challenge 2021 | 93.44 |
| 2 | Technical University of Munich | ASDNet at ActivityNet Challenge 2021 - Active Speaker Detection (AVA) | 91.85 |
| 3 | National University of Singapore | NUS-HLT Report for ActivityNet Challenge 2021 AVA (Speaker) | 90.82 |
The third AVA challenge ran as part of the International Challenge on Activity Recognition (ActivityNet) workshop at CVPR 2020. It included two tasks: action detection using the newly-released AVA-Kinetics dataset, and active speaker detection using AVA ActiveSpeaker.
The competition led to another significant improvement (+5.4 mAP) in action detection performance. A video presenting the results, including presentations by the winners of the action and active speaker tasks, can be found on YouTube as well as on bilibili.
Thanks as always to everyone who entered, and congratulations to the winners! Details of the winning methods are available in the reports below.
| Rank | Team Name | Entry Report | mAP |
| 1 | CUHK-SenseTime | Actor-Context-Actor Relation Network for Spatio-Temporal Action Localization | 39.62 |
| 2 | ByteDance & Shanghai Jiao Tong University | Multiple Attempts for AVA-Kinetics | 32.91 |
| 3 | Fujitsu | Multi-scale Spatiotemporal Features for Action Localization | 31.88 |
| Rank | Team Name | Entry Report | mAP |
| 1 | Universidad de los Andes | Active Speaker Detection | 86.68 |
The second AVA Challenge was held at the International Challenge on Activity Recognition (ActvityNet) workshop in conjunction with CVPR 2019. In addition to the spatio-tempral action recognition task, this year we introduced a new secondary task: active speaker detection.
Thank you to all who participated! A total of 30 teams entered, making a combined 79 submissions, achieving an impressive +13 mAP increase over last year's winner for the Actions task, and establishing a new, strong baseline for the Active Speaker task.
The winners of the challenge are below. You will be able to find the full leaderboard at the ActivityNet Challenge site.
| Rank | Team Name | Entry Report | mAP |
| 1 | FAIR | SlowFast Networks for Video Recognition | 34.25 |
| 2 | Machine Vision and Intelligence Group, Shanghai Jiao Tong University | Three Branches: Detecting Actions with Richer Features | 32.49 |
| 3 | Shanghai Jiao Tong University & ByteDance AI Lab | ByteDance AI Lab AVA Challenge 2019 Technical Report | 30.20 |
| Rank | Team Name | Entry Report | mAP |
| 1 | Naver Corporation | Naver at ActivityNet Challenge 2019 | 87.82 |
| 2 | University of Chinese Academy of Sciences | Multi Task Learning for Audio Visual Active Speaker Detection | 83.49 |
| - | Google baseline | AVA-ActiveSpeaker: An Audio-Visual Dataset for Active Speaker Detection | 82.08 |
The first AVA challenge was held in partnership with the ActivityNet workshop at CVPR 2018. Details on the specific task can be found here.
A total of 16 teams entered, making a combined 35 submissions, and achieving a +5.5 point increase in mAP over the baseline.
The winners of the challenge are below. You can find the full leaderboard at the ActivityNet Challenge site.
| Rank | Team Name | Entry Report | Task 1 mAP | Task 2 mAP |
| 1 | Tsinghua/Megvii | Human Centric Spatio-Temporal Action Localization | 21.08 | 20.99 |
| 2 | DeepMind | A Better Baseline for AVA | 21.03 | 21.03 |
| 3 | YH Technologies | YH Technologies at ActivityNet Challenge 2018 | 19.60 | 19.60 |