CALL FOR PARTICIPATION in the
2023 TREC VIDEO RETRIEVAL EVALUATION (TRECVID 2023)
February 2023 - December 2023
Conducted by the National Institute of Standards and Technology (NIST)
with additional funding from other US government agencies. Below you
can find an overview on the used datasets, tasks, and how to apply to
participate. All teams are encouraged to apply early to get access to
data and join the slack workspace of active teams.
Please consult the guidelines for each task for more details
including General Schedule
Application URL:
http://ir.nist.gov/tv-submit.open/application.html
Application deadline: June 1.
[apply early to get access to task discussions, participants mailing lists, and datasets]
Introduction:
The TREC Video Retrieval Evaluation series (trecvid.nist.gov) promotes
progress in content-based analysis of and retrieval from digital video
via open, metrics-based evaluation. TRECVID is a laboratory-style
evaluation that attempts to model real world situations or significant
component tasks involved in such situations. In its 23rd annual evaluation
cycle, TRECVID will evaluate participating systems on 5 different video
analysis and retrieval tasks (Adhoc video search, video to text captioning,
movie QA, medical video QA, and activities detection) using various types
of real world datasets. Please see below more details about the main
datasets & tasks to be used in 2023 across the 5 proposed tasks.
Data:
In TRECVID 2023 NIST will use at least the following data sets:
* Vimeo Creative Commons Collection (V3C)
The V3C is a large-scale video dataset that has been collected from high-quality
web videos with a time span over several years in order to represent true videos
in the wild. It consists of 28,450 videos with a duration of 3,801 hours in total.
In 2023, the V3C2 subcollection (1,300 hr and 1.4 million shots) will be utilized
as testing dataset, while V3C1 (1,000 hr and 1 million shots) previously adopted at
TRECVID from 2019-2021 as a development dataset. This new V3C2 subcollection is
planned to be adopted from 2022 to 2024.
* IACC.3
The IACC.3 was introduced in 2016 and consists of approximately 4600 Internet
Archive videos (144 GB, 600 h) with Creative Commons licenses in MPEG-4/H.264
format with duration ranging from 6.5 min to 9.5 min and a mean duration
of almost 7.8 min. Most videos will have some metadata provided by the
donor available e.g., title, keywords, and description. The
IACC.3 is provided as development dataset for teams.
* Kino lorber edu Movies
A set of 10 movies licensed from Kino Lorber Edu (https://www.kinolorberedu.com/)
will be available to support the deep video understanding (DVU) task. Five movies
will be assigned as part of the training dataset with annotations at the movie and
scene levels, while the other 5 movies will be employed as the testing dataset for
the DVU task. All movies are in English with duration between 1.5 - 2 hrs each.
Participants Will be able to download the whole original movies and use the data for
research only purpose within TRECVID tasks.
* Deep Video Understanding (DVU)
A set of 14 movies (total duration of 17.5 hr) with Creative Commons license previously
utilized at the ACM Multimedia Grand Challenges in 2020 and 2021 will be available as a
development dataset for the DVU task. The dataset contains movie-level and scene-level annotations.
The movies have been collected from public websites such as Vimeo and the Internet Archive.
In total, the 14 movies consist of 621 scenes, 1572 entities, 650 relationships, and 2491 interactions.
* TV_VTT
This dataset will support the training dataset for the Video-to-Text (VTT) task.
It contains short videos (ranging from 3 seconds to 10 seconds) from TRECVID VTT task
from 2016 to 2022. There are 12,870 videos with captions. Each video has between 2 and
5 captions, which have been written by dedicated annotators.
* MedVidQA Collections
The VCVAL (Video Corpus Visual Answer Localization) task is supported by MedVidQA
collections training dataset consisting of 3,010 human-annotated instructional questions
and visual answers from 900 health-related videos. In addition, an automatically created
HealthVidQA dataset consists of ~50 000 instructional questions and visual answers from
15,000 health-related videos. A validation dataset consisting of 50 questions and their
answer timestamps created from 25 medical instructional videos will also be available.
Finally, the testing dataset will contain 50 questions and their answer timestamps created
from 25 medical instructional videos.
The MIQG (Medical Instructional Question Generation) task is supported by a training dataset
consisting of 2710 question and visual segments, which are formulated from 800 medical instructional
videos from the MedVidQA collections. The provided validation dataset will contain 145 questions
and answers timestamps created from 49 medical instructional videos, while the test dataset will
contain 100 questions and answers timestamps created from 45 medical instructional videos.
* Gatwick and i-LIDS MCT airport surveillance video
The data consist of about 150 hours obtained from airport
surveillance video data (courtesy of the UK Home Office). The
Linguistic Data Consortium has provided event annotations for
the entire corpus. The corpus was divided into development and
evaluation subsets. Annotations for 2008 development and test
sets are available.
* MEVA dataset
The TRECVID ActEV 2023 Challenge is based on the Multiview Extended Video
with Activities (MEVA) Known Facility (KF) dataset. The large-scale MEVA
dataset is designed for activity detection in multi-camera environments.
It was created on the Intelligence Advanced Research Projects Activity (IARPA)
Deep Intermodal Video Analytics (DIVA) program to support DIVA performers and
the broader research community. You can download the public MEVA resources
(training video, training annotations and the test set) at mevadata.org
Tasks:
In TRECVID 2023 NIST will evaluate systems on the following tasks
using the [data] indicated:
* AVS: Ad-hoc Video Search (automatic, manually-assisted, relevance feedback) [V3C2]
The Ad-hoc search task started in TRECVID 2016 and will continue in 2023
to model the end user search use-case, who is looking for
segments of video containing persons, objects, activities, locations, etc.,
and combinations of the former. Given about 40 textual queries created at
NIST, return for each query all the shots which meet the video need expressed
by it, ranked in order of confidence. Although all evaluated submissions will be
for automatic runs, Interactive systems will have the opportunity to
participate in the Video Browser Showdown (VBS) in 2024 using the same testing data (V3C2).
* ActEV: Activities in Extended Video [MEVA]
ActEV is a series of evaluations to accelerate development of robust, multi-camera,
automatic activity detection algorithms for forensic and real-time alerting applications.
ActEV is an extension of the annual TRECVID Surveillance Event Detection (SED) evaluation
where systems will also detect, and track objects involved in the activities. Each evaluation
will challenge systems with new data, system requirements, and/or new activities.
* MedVidQA: Medical Video Question Answering [MedVidQA Collections]
Many people prefer instructional videos to teach or learn how to accomplish a particular task
with a series of step-by-step procedures in an effective and efficient manner. In a similar way,
medical instructional videos are more suitable and beneficial for delivering key information
through visual and verbal communication to consumers' healthcare questions that demand instruction.
With an aim to provide visual instructional answers to consumers' first aid, medical emergency, and
medical educational questions, this TRECVID NEW task on medical video question answering will
introduce a new challenge to foster research toward designing systems that can understand medical
videos to provide visual answers to natural language questions and equipped with the multimodal
capability to generate instructional questions from the medical video.
Following the success of the 1st MedVidQA shared task in the BioNLP workshop at ACL 2022, MedVidQA 2023 at
TRECVID expanded the tasks and introduced a new track considering language-video understanding and generation.
This track comprises two main tasks, Video Corpus Visual Answer Localization (VCVAL) and
Medical Instructional Question Generation (MIQG). For detailed information, please refer to
the task guidelines page.
* DVU: Deep Video Understanding [Kino lorber edu movies]
Deep video understanding is a difficult task which requires computer vision systems to develop
a deep analysis and understanding of the relationships between different entities in video,
and to use known information to reason about other, more hidden information.
The aim of the task is to push the limits of multimedia analysis techniques to
address analysing long duration videos holistically and extract useful knowledge to utilize it
in solving different kinds of queries. The knowledge in the target queries includes both visual
and non-visual elements. Participating systems should take into consideration all available
modalities (speech, image/video, and in some cases text).
The task for participating researchers will be: given a whole original movie (e.g 1.5 - 2hrs long),
image snapshots of main entities (persons, locations, and concepts) per movie, and ontology of
relationships, interactions, locations, and sentiments used to annotate each movie at global movie-level
(relationships between entities) as well as on fine-grained scene-level (scene sentiment, interactions
between characters, and locations of scenes), systems are expected to generate a knowledge-base of
the main actors and their relations (such as family, work, social, etc) over the whole movie, and
of interactions between them over the scene level. This representation can be used to answer a set of
queries on the movie-level and/or scene-level per movie. The task will support two tracks (subtasks)
where teams can join one or both tracks. Movie track where participants are asked queries on the whole
movie level, and Scene track where Queries are targeted towards specific movie scenes. New this year,
is a subtask where systems can also submit results against the same queries but modified testing dataset
after introducing some natural corruptions and perturbations to simulate real world noise datasets.
* VTT: Video to Text Description [V3C3]
Automatic annotation of videos using natural language text descriptions has been a long-standing goal
of computer vision. The task involves understanding of many concepts such as objects, actions,
scenes, person-object relations, temporal order of events and many others. In recent years there have
been major advances in computer vision techniques which enabled researchers to start practically to
work on solving such problem. Given a set of short video clips, systems are asked to work and
submit results for a main task: The "Description Generation" task requires systems to automatically
generate a text description (1 sentence) for each video clip based on who is doing what, where and when.
The other subtask proposed this year is to generate text descriptions on the same testing dataset but
after introducing some natural corruptions and perturbations to simulate real world noise datasets.
In addition to the data, TRECVID will provide uniform scoring procedures, and a forum for organizations
interested in comparing their approaches and results.
Participants will be encouraged to share resources and intermediate system outputs to lower entry barriers
and enable analysis of various components' contributions and interactions.
***************************************************
* You are invited to participate in TRECVID 2023 *
***************************************************
The evaluation is defined by the Guidelines. A draft version is
available and further feedback input from the participants are welcomed till April,2023.
You should read the guidelines carefully before applying to participate in one or more tasks:
Guidelines
Please note
1) Dissemination of TRECVID work and results other than in the
(publicly available) conference proceedings is welcomed, but the
conditions of participation specifically preclude any advertising
claims based on TRECVID results.
2) All system output and results submitted to NIST are published in
the Proceedings or on the public portions of TRECVID web site archive.
3) The workshop is open to participating groups that submit
results for at least one task, to selected government personnel
from sponsoring agencies, data donors, and interested researchers
who may never participated Before and would like to know more about TRECVID.
4) Each participating group is required to submit before the
workshop a notebook paper describing their experiments and results.
This is true even for groups who may not be able to attend the
workshop.
5) It is the responsibility of each team contact to make sure that
information distributed via the call for participation and the
[email protected] email list is disseminated to all team members with
a need to know. This includes information about deadlines and
restrictions on use of data.
6) By applying to participate you indicate your acceptance of the
above conditions and obligations.
There is a tentative schedule for the tasks included in the Guidelines
webpage: Schedule
Workshop format
The workshop format as being in-person, hybrid, Or virtual in 2023 is still
something to be decided. Details will be provided to participants as soon as available.
The TRECVID workshop is used as a forum both for presentation of
results (including failure analyses and system comparisons), and for
more lengthy system presentations describing retrieval techniques
used, experiments run using the data, and other issues of interest to
researchers in information retrieval and computer vision. As there is
a limited amount of time for these presentations, the evaluation coordinators
and NIST will determine which groups are asked to speak and which groups will
present in a poster session. Groups that are interested in having a
speaking slot during the workshop will be asked to submit a short
abstract before the workshop describing the experiments they
performed. Speakers will be selected based on these abstracts.
How to respond to this call
Organizations wishing to participate in TRECVID 2023 must respond
to this call for participation by submitting an on-line application by
the latest 1 June (the earlier the better). Only ONE APPLICATION PER TEAM
please, regardless of how many organizations the team comprises.
*PLEASE* only apply if you are able and fully intend to complete the
work for at least one task. Taking the data but not submitting any
runs threatens the continued operation of the workshop and the
availability of data for the entire community.
Here is the application URL:
http://ir.nist.gov/tv-submit.open/application.html
You will receive an immediate automatic response when your application
is received. NIST will respond with more detail to all applications submitted
before the end of March. At that point you'll be given the active participant's
userid and password, be subscribed to the tv23.list email discussion list,
and can participate in finalizing the guidelines as well as sign up to get the data,
which is controlled by separate passwords. All active teams will also be added
to a slack workspace to encourage more communication and facilitate announcements.
TRECVID 2023 email discussion list
The tv23.list email discussion list ([email protected]) will serve as
the main forum for discussion and for dissemination information about
TRECVID 2023. It is each participant's responsibility to monitor the
tv23.list postings. It accepts postings only from the email addresses
used to subscribe to it. An archive of past postings is available using the active
participant's userid/password.
Questions ?
Any administrative questions about conference participation,
application format/content, subscriptions to the tv23.list,
etc. should be sent to george.awad at nist.gov.
Best regards,
TRECVID 2023 organizers team
