The main goal of the TREC Video Retrieval Evaluation (TRECVID) is to promote 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.
Up until 2010, TRECVID used test data from a small number of known professional sources - broadcast news organizations, TV program producers, and surveillance systems - that imposed limits on program style, content, production qualities, language, etc. In 2003 - 2006 TRECVID supported experiments in automatic segmentation, indexing, and content-based retrieval of digital video using broadcast news in English, Arabic, and Chinese. TRECVID also completed two years of pilot studies on exploitation of unedited video rushes provided by the BBC. In 2007 - 2009 TRECVID provided participants with cultural, news magazine, documentary, and education programming supplied by the Netherlands Institute for Sound and Vision. Tasks using this video included segmentation, search, feature extraction, and copy detection. Systems were tested in rushes video summarization using the BBC rushes. Surveillance event detection was evaluated using airport surveillance video provided by the UK Home Office. Many resources created by NIST and the TRECVID community are available for continued research on this data independent of TRECVID. See the Past data section of the TRECVID website for pointers.
In 2010 TRECVID confronted known-item search and semantic indexing systems with a new set of Internet videos (referred to in what follows as IACC) characterized by a high degree of diversity in creator, content, style, production qualities, original collection device/encoding, language, etc - as is common in much "Web video". The collection also has associated keywords and descriptions provided by the video donor. The videos are available under Creative Commons licenses from the Internet Archive. The only selection criteria imposed by TRECVID beyond the Creative Commons licensing is one of video duration - they are short (less than 6 min). In addition to the IACC data set, NIST began developing an Internet multimedia test collection (HAVIC) with the Linguistic Data Consortium and used it in growing amounts (up to 4000 h) in TRECVID 2010-present. The airport surveillance video, introduced in TRECVID 2009, has been reused each year since.
New in 2013 will be video provided by the BBC. Programming from their long-running EastEnders series will be used in the instance search task. An additional 600 h of Internet Archive video available under Creative Commons licensing for research (IACC.2) will be used for the semantic indexing task.
In TRECVID 2013 NIST will evaluate systems on the following tasks using the [data] indicated:
A number of datasets are available for use in TRECVID 2013 and are described below.
Three datasets (A,B,C) - totaling approximately 7300 Internet Archive videos (144 GB, 600 h) with Creative Commons licenses in MPEG-4/H.264 format with duration ranging from 10 s to 6.4 min and a mean duration of almost 5 min. Most videos will have some metadata provided by the donor available e.g., title, keywords, and description
Data use agreements and Distribution: Download for active participants from NIST/mirror servers. See Data use agreements
Master shot reference: Will be available to active participants by download from the active participant's area of the TRECVID website.
Master I-Frames: Will be extracted by NIST using ffmpeg for all videos in the IACC.2.A collection and made available to active participants by download from the active participant's area of the TRECVID website.
Automatic speech recognition (for English): We hope to be able to provide this from the same source as in previous years for the IACC.1 collection.
Three datasets (A,B,C) - totaling approximately 8000 Internet Archive videos (160 GB, 600 h) with Creative Commons licenses in MPEG-4/H.264 format with duration between 10s and 3.5 min. Most videos will have some metadata provided by the donor available e.g., title, keywords, and description
Data use agreements and Distribution: Available by download from the Internet Archive. See TRECVID Past Data page. Or download from the copy on the Dublin City University server, but use the collection.xml files (see TRECVID past data page) for instructions on how to check the current availability of each file.
Master shot reference: Available by download from the TRECVID Past Data page
Automatic speech recognition (for English): Available by download from the TRECVID Past Data page
Approximately 3200 Internet Archive videos (50 GB<, 200 h) with Creative Commons licenses in MPEG-4/H.264 format with durations between 3.6 and 4.1 min Most videos will have some metadata provided by the donor available e.g., title, keywords, and description
Data use agreements and Distribution: Available by download from the Internet Archive. See TRECVID Past Data page.
Master shot reference: Available by download from the TRECVID Past Data page
Common feature annotation: Available by download from the TRECVID Past Data page
Automatic speech recognition (for English): Available by download from the TRECVID Past Data page
The data consist of about 150 h of 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.
Data use agreements and Distribution:
Development data annotations: available by download.
Approximately 244 video files (totally 300 GB, 464 h) with associated metadata, each containing a week's worth of BBC EastEnders programs in MPEG-4/H.264 format.
Data use agreements and Distribution: Download and fill out the data permission agreement from the active participants' area of the TRECVID website. After the agreement has been processed by NIST and the BBC, the applicant will be contacted by Dublin City Univerisity with instructions on how to download from their servers. See Data use agreements
Master shot reference: Will be available to active participants by download from TRECVID 2013 active participant's area.
Automatic speech recognition (for English): Will be available to active participants by download from Dublin City University.
HAVIC is a large collection of Internet multimedia constructed by the Linguistic Data Consortium and NIST. Participants will receive training corpora, event training resources, and two development test collections. Participants will also receive the ~4,000 h MED Progress evaluation collection which was used for MED '12 and will continue to be used through MED '15 as the test collection.
Data use agreements and Distribution: Data licensing and distribution will be handled by the Linguistic Data Consortium. See the Multimedia Event Detection task webpage for details.
In order to be eligible to receive the data, you must have have applied for participation in TRECVID. Your application will be acknowledged by NIST with a team ID, and active participant's password, and information about how to obtain the data.
As Subject: "TRECVID data request"
In the body: your name
your short team ID (given when you applied to participate)
the kinds of data you will be using - one or more of the following:
Gatwick (2008), IACC.2, and/or BBC EastEnders
You will receive instructions on how to download the data.
Please ask only for the test data (and optional development data) required for the task(s) you apply to participate in and intend to complete.
Requests are handled in the order they are received. Please allow 3 business days for NIST to respond to your request with the access codes you need to download the data using the information about data servers in the the active participant's area.
This task will be coordinated by Georges Quénot - with Han Dong from the Laboratoire d'Informatique de Grenoble and Stéphane Ayache from the Laboratoire d'Informatique Fondamentale de Marseille using support from the Quaero Programme and in collaboration with NIST. Cees Snoek and Xirong Li from University of Amsterdam will participate in the selection of the concept pairs for the paired-concept task.
Automatic assignment of semantic tags representing visual or multimodal concepts (previously "high-level features") to video segments can be fundamental technology for filtering, categorization, browsing, search, and other video exploitation. New technical issues to be addressed include methods needed/possible as collection size and diversity increase, when the number of concepts increases, and when concepts are related by an ontology. In 2013 the task will again support experiments in two areas introduced in 2012:
NEW! In addition, as an experiment, there will be an optional
system output for each shot returned, which holds the component
concept number (listed below) of the concept which appears first in
the shot; a value of 0 for this output will indicate that the two
concepts first appear simultaneously. Note that this does not change
the rule that both concepts must occur together in at least one frame
of the shot. The submission format will be updated by 1. May to reflect this
addition. Results for this optional output will be returned and
reported at the workshop, but will not affect the primary measures for
the task.
[911] Telephones (117) + Girl (54)
[912] Kitchen (72) + Boy (16)
[913] Flags (261) + Boat_Ship (15)
[914] Boat_Ship (15) + Bridges (17)
[915] Quadruped (392) + Hand (59)
[916] Motorcycle (80) + Bus (19)
[917] Chair (25) + George_[W_]Bush(274)
[918] Flowers (53) + Animal (6)
[919] Explosion_Fire (49) + Dancing (38)
[920] Government-Leader (56) + Flags (261)
In addition, the semantic indexing task will introduce two new opportunities for experimentation:
Main: Given the test collection (IACC.2.A), master shot reference, and single concept definitions, return for each target concept a list of at most 2000 shot IDs from the test collection ranked according to their likelihood of containing the target.
Localization subtask: For each concept from the list of 10 designated for localization, for each shot of the top-ranked 1000 returned in a main task run, for each I-Frame within the shot that contains the target, return the x,y coordinates of the upper left and lower right vertices of a bounding rectangle which contains all of the target concept and as little more as possible. Systems may find more than one instance of a concept per I-Frame and then may include more than one bounding box for that I-Frame, but only one will be used in the judging since the ground truth will contain only 1 per judged I-Frame, one chosen by the NIST assessor, at least in this first round.
Progress: The progress task is just the main task run additionally, independently on the progress data sets for 2013: just IACC.2.B and just IACC.2.C. (No localization)
Paired: Given the test collection (IACC.2.A), master shot reference, and concept-pair definitions, return for each target concept pair a list of at most 2000 shot IDs from the test collection ranked according to their likelihood of containing the target. In 2013 each participant in the paired concept task must submit a baseline run which just combines for each pair the output of group's two independent single-concept detectors. Optionally, participants may submit information for all pairs, indicating the temporal sequence in which the two concepts occur. Note this does not change the requirement that both concepts must occur within at least one frame in the shot in order for the shot to be considered as containing the concept.
The current test data set (IACC.2.A) will be 200 h drawn from the IACC.2 collection using videos with durations between 10 s and 6 min.
The progress test data sets (IACC.2.B-C) will be 2 additional non-overlapping collections of 200 h (IACC.2.B and IACC.2.C) each drawn randomly from the IACC.2 collection.
The development data set combines the development and test data sets of the 2010 and 2011 issues of the task, IACC.1.tv10.training, IACC.1.A, IACC.1.B, and IACC.1.C each containing about 200 h drawn from the IACC.1 collection using videos with durations ranging from 10s to just longer than 3.5 min. These datasets can be downloaded from the Internet Archive using information available on the TRECVID "past data" webpage.
500 concepts were selected for the TRECVID 2011 semantic indexing task. In making this selection, the organizers drew from the 130 used in TRECVID 2010, the 374 selected by CU/Vireo for which there exist annotations on TRECVID 2005 data, and some from the LSCOM ontology. From these 500 concepts, 346 concepts were selected for the full task in 2011 as those for which there exist at least 4 positive samples in the final annotation. A spreadsheet of the concepts is available here with complete definitions and an alignment with CU-VIREO374 where appropriate. [Don't be confused by the multiple numberings in the spreadsheet - use the TV13-15 IDs in the concept lists below under "Submissions".] For 2013 the same list of 500 concepts has been used as a starting point for selecting the 60 single concepts for which participants must submit results in the main task and the 10 concept pairs in the paired concept task. The concepts (number to be determined) for localization will be a subset of the main task concepts - perhaps about 10.
The organizers have provided again a set of relations between the concepts. There are two types of relations: A implies B and A excludes B. Relations that can be derived by transitivity will not be included. Participants are free to use the relations or not and submissions are not required to comply with them.
It is expected that advanced methods will use the annotations of non-evaluated concepts and the ontology relations to improve the detection of the evaluated concepts. The use of the additional annotations and of ontology relations is optional and comparison between methods that use them and methods that do not is encouraged.
Four types of submissions will be considered:
P l e a s e n o t e t h e s e r e s t r i c t i o n s and this information on training types. The submission types (main and pair) are orthogonal to the training types (A, B, C ...).
Each team may submit a maximum of 4 prioritized main runs with 2 additional if they are of the "no annotation" training type and the others are not. One localization run may be submitted with each main submission. Each team may also submit up to 2 "pair" runs. Each team may submit up to 2 progress runs on each of the 2 progress datasets. The submission formats are described below.
Please note: Only submissions which are valid when checked against the supplied DTDs will be accepted. You must check your submission before submitting it. NIST reserves the right to reject any submission which does not parse correctly against the provided DTD(s). Various checkers exist, e.g., Xerces-J: java sax.SAXCount -v YourSubmision.xml.
Concept# | File# Frame# | | | UpperLeftX | | | | UpperLeftY | | | | | LowerRightX | | | | | | LowerRightY | | | | | | | xxx xxxxx xxxx xxx xxx xxx xxx
A subset of the submitted concept results (at least 20), to be announced only after the submission date, will be evaluated by assessors at NIST or at LIG using pooling and sampling.
Please note that NIST uses a number of rules in manual assessment of system output.
Measures (indexing):
Measures (localization): Temporal and spatial localization will be
evaluated using precision and recall based on the judged items at two
levels - the frame and the pixel, respectively. NIST will then
calculate an average for each of these values for each concept and for
each run.
For each shot that is judged to contain a concept, a
subset of the shot's I-Frames will be viewed and annotated to locate
the pixels representing the concept. The set of annotated I-Frames
will then be used to evaluate the localization for the I-Frames
submitted by the systems.
For each run, a total elapsed time in seconds will be reported.
How can we encourage use of the (interframe) video context in improving localization over that done for isolated images? For example, assuming interframe context decreases with growing interframe time interval, should we make the sample of I-Frames we decide to judge for a given shot-concept "contiguous" in time?
Detecting human behaviors efficiently in vast amounts surveillance video is fundamental technology for a variety of higher-level applications of critical importance to public safety and security. The use case addressed by this task is the retrospective exploration of surveillance video archives using a system designed to support the optimal division of labor between a human user and the software - an interactive system.
Given a collection of surveillance data files (e.g. that from an airport, or commercial establishment) for preprocessing, at test time take a small set of topics (search requests for known events) and for each return the elapsed search time and a list of video segments within the surveillance data files, ranked by likelihood of meeting the need described in the topic. Each search for an event by a searcher can take no more than 25 elapsed minutes, measured from the time the searcher is given the event to look for until the time the result set is considered final.
The test data will be the same as was used in the SED task in 2012.
Submissions will follow the same format and procedure as in the SED 2012 task. The number of submissions allowed will be determined by the time the Guidelines are final. Participants must submit at least one interactive run. An automatic version of each interactive run for comparison may also be submitted.
It is assumed the user(s) are system experts and no attempt will be made to separate the contribution of the user and the system. The results for each system+user will be evaluated by NIST as to effectiveness - using standard search measures (e.g., probability of missed detection/false alarm, precision, recall, average precision)- self-reported speed, and user satisfaction (for interactive runs).
An important need in many situations involving video collections (archive video search/reuse, personal video organization/search, surveillance, law enforcement, protection of brand/logo use) is to find more video segments of a certain specific person, object, or place, given a visual example.
In 2013 NIST will create about 2 dozen topics. Automatic and interactive systems will use the same set. New data will be available from the BBC EastEnders television series, prepared with major help from several participants in the AXES project (access to audiovisual archives), a four-year FP7 framework research project to develop tools that provide various types of users with new and engaging ways to interact with audiovisual libraries.
Given a collection of test video, a master shot reference, and a collection of topics (queries) that delimit a person, object, or place entity in some example video, locate for each topic up to the 1000 shots most likely to contain a recognizable instance of the entity. Interactive runs will likely return many fewer than 1000 shots.
Development data: A very small sample of the BBC Eastenders test data will be available from Dublin City University. No actual development data will be supplied.
Test data: The test data for 2013 will be BBC EastEnders video in MPEG-4 format. The example images in the topics will be in bmp format. See above for information on how to get a copy of the test data.
Topics: Each topic will consist of a set of 4 example frame images (bmp) drawn from test videos containing the item of interest. The shots from which example images are drawn for a given concept, will be filtered by NIST from system submissions for that concept before evaluation. For each frame image a binary mask of the region of interest will be provided. Each topic will also include an indication of the target type taken from this set of strings {PERSON, LOCATION, OBJECT}
Auxilliary data: Participants are allowed to use various publicly available EastEnders resources as long as they carefully note the use of each such resource by name in their workshop notebook papers. They are strongly encouraged to share information about the existence of such resources with other participants via the tv13.list as soon as they discover them.
Each team may submit a maximum of 4 prioritized runs. All runs will be evaluated but not all may be included in the pools for judgment. Submissions will be identified as either fully automatic or interactive. Interactive runs will be limited to 15 elapsed minutes per search.
Please note: Only submissions which are valid when checked against the supplied DTDs will be accepted. You must check your submission before submitting it. NIST reserves the right to reject any submission which does not parse correctly against the provided DTD(s). Various checkers exist, e.g., Xerces-J: java sax.SAXCount -v YourSubmision.xml.
Here for download (though they may not display properly) is the DTD for search results of one run, the one for results from multiple runs, and a small example of what a site would send to NIST for evaluation. Please check your submission to see that it is well-formed
You may submit all your runs in one or multiple files as long as you do not break a run across files. EACH file you submit should begin, as in the example submission, with the DOCTYPE statement and a videoSearchResults element even if only one run is included.
Submissions must be transmitted to NIST via this webpage.
This task will be treated as a form of search and will accordingly be evaluated with average precision for each topic in each run and per-run mean average precision over all topics. Speed will also be measured.
Video is becoming a new means of documenting everything from recipes to how to change a tire of a car. Ever expanding multimedia video content necessitates development of new technologies for retrieving relevant videos based solely on the audio and visual content of the video clip.
Participants are tasked with building an automated system that can determine whether an event is present anywhere in a video clip using the content of the video clip only. The system inputs a set of "search" videos and an "event kit" (text and video describing the event). The system computes an "event score" (that gives the strength of evidence of the event) and an optional "recounting" of the event in the each search video in the input set.
The 2013 evaluation will consist of 20 "pre-specified" and 20 new "ad-hoc" event kits containing 100, 10 or 0 example event videos. Developers will receive the pre-specified event kits in advance of building metadata generator (which extracts and saves a representation for each video's content). Once the metadata generator has been locked and the metadata is created for the search video set, developers will process the pre-specified and ad-hoc events.
Participants may choose to either run their system on a set of 98,000 search videos or a subset containing approximately 32,000 video clips.
All participants must submit the results from their system that: (1) processes all 20 pre-specified, 100-example event kits and (2) processes either the evaluation search video set or its specified subset.
Participants choosing to process the 20 ad-hoc events must submit results from their system that: (1) processes all 20 ad-hoc, 100-example event kits and (2) processes either the evaluation search video set or its specified subset.
Participants may optionally submit results from their system that use fewer event kit video examples (i.e., 10 and 0 examples). Participants may also optionally submit partial results of their primary system (using 100 event examples) for one or more pre-defined conditons:
Participants will be provided new data divisions to facilitate cross-team comparison of results. The four data divisions will include research data, training data, an evaluation search video set, and two test search video sets.
Research data: Participants will be provided data and research event kits to develop algorithms and systems to design their video content metadata representation. This data may be augmented, annotated or altered to support each participant's research and development of their algorithms. However, none of this data may be used directly for training pre-specified or ad-hoc event classifiers to be used for the system.
Training data: This data consists of the event kit example videos and background training videos. The event kit text will identify required and supporting observable evidence of an event. This information can be used to inform the metadata representation and to design algorithms. This data is the only data to be used to train event classifiers. However, this data must NOT be extended, annotated, nor altered.
Test search video sets: Participants will be provided two search video sets for testing and publication of results. These videos cannot be used in event training nor defining/implementing the metadata representation or extraction. This data must NOT be extended, annotated nor altered.
Evaluation Search Set: This data will be used for blind testing of participants' system and will be the Progress Set (4,000 h of search videos used in MED'12) and a pre-defined Progress Subset (a 1,300 h subset of the Progress Set). This data must be kept "blind" and teams must not view or analyze any properties of the video set.
Please refer to 2013 MED evaluation plan for instructions on the submission process.
Event Detection performance will be assessed using a Mean Average Precision as the primary evaluation measure. Detection threshold setting will be evaluated using a to-be-determined measure. NIST will provide the ground truth and evaluation tools. All participants will report their results in their TRECVID workshop notebook papers.
Please refer to the 2013 MED evaluation plan for instructions on the submission process.
A MED system not only detects events in video clips but also recounts the evidences used to identify the event. These recountings help the user quickly and accurately locate their event of interest within the clips detected by the MED system.
The system task is to provide a recounting of the important evidence that a video clip contains an instance of an event of interest. For each piece of evidence, the recounting must include both text summarizing the evidence and a list of one or more spatiotemporal pointers (in a text format) locating the evidence within the clip. NIST will provide a DTD for the XML format.
MER participants are the participants in the MED evaluation whose MED submission also includes a recounting for each clip. If no evidence of the event is found by the MED system, the recounting will include no evidence. The recountings to be evaluated will be selected from
Additional details of the evaluation procedures are in the 2013 MER Evaluation Specification Documents, which will appear in the near future and before the Guidelines are final.
See MED data section.
MER evaluates the recountings submitted with each participant’s MED results.
Submissions will be XML-compliant plain-ASCII text. A MER DTD and triage workstation will be made available to all participants.
At most one submission per participating MED team will be accepted. Due to limits on manual judging resources, NIST may need to restrict the number of submissions that are judged. If so, this will be based on the performance of the associated MED system and order of MER submission (first come, first served).
Please refer to the 2013 MER Evaluation Specification Document, when complete, for instructions on the submission process.
The system's recountings will be evaluated by a panel of judges. Each judge will be provided with a "triage workstation," which will take as input the event kit (or query for the event of interest), the system's recountings that are being assessed, and the clips associated with each of those recountings. NIST will choose a subset of clips for which recountings will be judged for selected events. All submissions will be judged on the same set of clips and events. For each such event, the judges will try to (as rapidly as possible) use the recountings to find clips that contain an instance of the event of interest.
In the triage workstation, the judge will be shown all the pieces of evidence in the recounting, both the textual and the audiovisual via the spatiotemporal pointers. Each spatiotemporal pointer will have a temporal part that specifies the period of time in the clip (the snippet or excerpt) where the piece of evidence occurs. For pieces of evidence that are visual, the spatiotemporal pointer will also include a bounding box within the video frame at the beginning and at the end of the snippet. After reading the recounting and hearing/viewing all the snippets for it, the judge will indicate whether he/she believes the associated clip contains an instance of the event of interest; the judge will have three choices (1) does contain, (2) does not contain, or (3) cannot readily tell whether it contains an instance of the event of interest (the difficulty could lie in the recounting or in the event kit).
For each submission, for each event, NIST will determine how quickly and accurately the judge is able to use the recounting to determine if the clip contains the event. Submissions whose recountings enable the judges to perform that task the most rapidly and accurately will be considered the best.
All participants will report their results in their TRECVID workshop notebook papers.
The latest details will be available soon and before the Guidelines are final on the MER webpage.
The following are the target dates for 2013:
Here is a list of work items that must be completed before the guidelines are considered to be final..
Pat Doe <[email protected]>
Once subscribed, you can post to this list by sending you thoughts as email to [email protected], where they will be sent out to EVERYONE subscribed to the list, i.e., all the other active participants.
Last
updated: