UVO

Unidentified Video Objects

A Benchmark for Dense, Open-World Segmentation

What's new:

2022 Information:

July 16th: UVO v1.0 is released! With 2X dense annotations and COCO category labels on dense set!

July 9th: We are hosting the 2nd challenge and workshop jointly with MOTComplex at ECCV 2022!

2021 Information:

October 18: Workshops recordings are available in Workshop Program section.

October 16: Winner announced: team Alpes_Runner for both Track 1 and Track 2! Congrats to team members: Yuming Du, Wen Guo, Yang Xiao, Vincent Lepetit!

October 7: Submit a technical report (1-4 pages) to uvo.dataset@gmail.com by October 13th, AoE, to be considered for awards! For details, please refer to Challenge Intro Section.

October 4: Given requests from participants and a delay in releasing the dataset, we are extending the Test server to be opened until October 8th (Friday) AoE.

September 24: Our ICCV21 workshop will be hosted on October 16th from 8AM to 12PM (EDT, GMT-4).

September 10: We received common requests for deleted videos and pre-processed videos to match the annotations. Due to Policy constraints we are not able to directly provide and host videos. However, we thank Tarun Kalluri who is also working on the dataset and willing to share his pre-processed videos that match baseline performances: Google Drive Link. You can use our provided video2frames.py script to split videos into frames. This will be a transient solution while we look for best way to host dataset videos. Thanks for the support from the community!

August 27: UVO v0.5 test frames/ videos released, example submissions released, example evaluation script released: Google Drive Link

August 27: Evaluation servers up on EvalAI, Track 1 Server, Track 2 Server

August 26: Call for Participants to Challenge@ICCV2021

August 13: UVO v0.5 released: Google Drive Link

July 12: Playset released: Google Drive Link

April 16: Website launched

April 12: UVO paper is on ArXiv: https://arxiv.org/abs/2104.04691

UVO Highlights

  • High quality instance masks densely annotated at 30 fps on 1024 YouTube videos and 1fps on 10337 videos from Kinetics dataset

  • Open-world: annotating all objects in each video, 13.5 objects per video on average

  • Diverse object categories: 57% of objects are not covered by COCO categories

A sneak peak of UVO

People

Weiyao Wang

Facebook AI Research

Matt Feiszli

Facebook AI Research

Heng Wang

Facebook AI Research

Du Tran

Facebook AI Research

UVO Paper

Contact us:

See authors of the paper