DeeperAction Workshop @ ICCV2021

Challenge and Workshop on Localized and Detailed Understanding of Human Actions in Videos
Monday 11th October 2021

NEWS

• 2021-10-19
  
  • Slides are now available at Program Schedule.
• 2021-10-05
  
  • Winners are announced and reports can now be downloaded in Challenge Results.
  • Details about our workshop schedule have been updated in Program Schedule.
• 2021-09-17
  
  • The leader board is released. Thanks for participation in our challenge. In summary, we receive 42 valid submissions for Track 1, 42 submissions for Track 2, 58 submissions for Track 3. Please check the competition pages for detailed information.
• 2021-09-10
  
  • Due to a problem with the codalab server which causes uploading failure, we decide to postpone the test phase and winner annonuncement for 2 days.
• 2021-07-07
  
  • Participants can now register on MultiSports (Track 2) - test. Please submit results for evaluation on this website.
• 2021-06-01
  
  • We have published competitions for FineAction (Track 1), MultiSports (Track 2) and Kinetics-TPS (Track 3). Competitions start at June 1, 2021, midnight UTC and finish at Sept. 10, 2021, 11:59 p.m. UTC.
  • Data for training and validation is now available on the competition pages above.
• 2021-05-26
  
  • Paper on FineAction (dataset for Track 1) can now be found here.
    
  • Paper on MultiSports (dataset for Track 2) can now be found here.

Attending the Workshop

ICCV will once again be virtual this year and the workshop will be held on Monday 11th October 2021 as a half day event. Please refer to this page for more information.

Aims and Scope

DeeperAction aims to advance the area of human action understanding with a shift from traditional action recognition to deeper understanding tasks of action, with a focus on localized and detailed understanding of human action from videos in the wild. Specifically, we benchmark three related tasks on localized and detailed action understanding by introducing newly-annotated and high-quality datasets, and organize the action understanding challenge on these benchmarks.

Temporal action Localization

Detecting all segments of containing actions of interest and recognizing their categories from a long video sequence.

Spatio-temporal action detection

Localizing all action instances with spatio-temporal tubes and recognizing their labels from untrimmed and multi-person videos.

Part-level action parsing

Decomposing the action instance into a human part graph and detecting action labels for all human parts an as well the whole human.