TRECVid Event Detection Evaluation Plan

2009 TRECVid Event Detection Evaluation Plan

1. Overview

This document presents the evaluation plan for event detection in surveillance video for TRECVid 2009. The goal of the evaluation will be to build and evaluate systems that can detect instances of a variety of observable events in the airport surveillance domain. The video source data to be used is a ~100-hour corpus of video surveillance data collected by the UK Home Office at the London Gatwick International Airport.

Two event detection tasks will be supported: a retrospective event detection task run with complete reference annotations, and a “freestyle” experimental analysis track to permit participants to explore their own ideas with regard to the airport surveillance domain.

The following topics are discussed below:

· Video source data

· Evaluation tasks

· Evaluation measures

· Evaluation Infrastructure

· Event definitions

· Schedule

2. Video Source Data

The source data will consist of at least 100 hours (10 days * 2 hours/day * 5 cameras), obtained from Gatwick Airport surveillance video data (courtesy of the UK Home Office). The Linguistic Data Consortium will provide event annotations for the entire corpus.

The corpus will be divided into development and evaluation subsets. The exact partitioning of the data is to be determined.

Developers may use the devset in any manner to build their systems, including activities such as dividing it into internal test sets, jackknifed training, etc. During the summer months, NIST will conduct a dry run evaluation using the devset as the video source. While testing on the development data is a non-blind test, the purpose of the dry run (to test the evaluation infrastructure) is most easily accomplished using the devset.

We will release the full corpus (devset + evalset) early in the evaluation cycle to give people the opportunity to preprocess the full corpus throughout the year. The evaluation set must not be inspected or mined for information until after the evalset annotations are released. The evalset restriction applies to both evaluation tasks. However, participants can run feature extraction programs on the evalset to prepare for the formal evaluation.

Allowable side information (i.e., “contextual” information) will include resources posted on the TRECVid Event Detection website as well as any annotations constructed by developers based on the devset. Participants may share devset annotations. No annotation of the evalset is permitted prior to the evaluation submission deadline.

3. Evaluation tasks

This proposal includes the following evaluation tasks:

Retrospective Event Detection: The task is to detect observations of events based on the event definition. Systems may process the full corpus using multiple passes prior to outputting a list of putative events observations. The primary condition for this task will be single-camera input (i.e., the camera views are processed independently). Multiple-camera input may optionally be run as an additional contrastive condition.

"Free-Style" Exercise. The purpose of this exercise is to support innovation and exploration of event detection in ways not anticipated by the retrospective task. Freestyle participants must define tasks that are pertinent to the airport video surveillance domain and that can implemented on this data set. Freestyle submissions must include rationale, clear definitions of the task, performance measures, reference annotations and a baseline system implementation.

4. Evaluation Infrastructure

Systems will be evaluated on how well they can detect event occurrences in the evaluation corpus. The determination of correct detection will be based solely on the temporal similarity between the annotated reference event observations and the system-detected event observations.

System detection performance is measured as a tradeoff between two error types: missed detections (MD) and false alarms (FA). The two error types will be combined into a single error measure using the Detection Cost Rate (DCR) model, which is a linear combination of the two errors. The DCR model distills the needs of a hypothetical application into a set of predefined constant parameters that include the event priors and weights for each error type. While the chosen constants have been motivated by discussions with the research and user communities, the single operation point characterized by the DCR model is a small window into the performance of an event detection system. In addition to DCR measures, Detection Error Tradeoff (DET) curves will be produced to graphically depict the tradeoff of the two error types over a wide range of operational points. The DCR model and the DET curve are related: the DCR model defines an optimal point along the DET curve.

The rest of this section defines the system output, followed by the three steps of the evaluation process: temporal alignment, Decision Error Tradeoff (DET) curve production, and DCR computations.

4.1. System Outputs

Systems will record observations of events in a VIPER-formatted XML file as described in the “TRECVid 2008 Event Detection: ViPER XML Representation of Events” document. Each event observation generated by a system will include the following items:

· Start frame: The frame number indicating the beginning of the observation (the first frame in the video source file is frame #1.)

· End frame: The frame number indicating the last frame of the observation.

· Decision score: A numeric score indicating how likely the event observation exists with more positive values indicating more likely observations.

· Actual Decision: A Boolean value indicating whether or not the event observation should be counted for the primary metric computation.

The decision scores and actual decisions permit performance assessment over a wide range of operating points. The decision scores provide the information needed to construct the DET curve. In order to construct a fuller DET curve, a system must over-generate putative observations far beyond the optimal point for the system’s best DCR value. The actual decisions provide the mechanism for the system to indicate which putative observations to include in the DCR calculation: i.e., the putative decisions with a true actual decision.

Systems must ensure their decision scores have the following two characteristics: first, the values must form a non-uniform density function so that the relative evidential strength between two putative terms is discernable. Second, the density function must be consistent across events for a single system so that event-averaged measures using decision scores are meaningful.

Since the decision scores are consistent across events, the system must use a single threshold for differentiating true and false actual decisions.

4.2. Event Alignment

Event observations can occur at any time and for any duration. Therefore, in order to compare the output of a system to the reference annotations, an optimal one-to-one mapping is needed between the system and reference observations. The mapping is required because there is no pre-defined segmentation in the video. The alignment will be performed using the Hungarian Solution to the Bipartite Graph [1] matching problem by modeling event observations as nodes in the bipartite graph. The system observations are represented as one set of nodes, and the reference observations are represented as a second set of nodes. The kernel formulas below assume the mapping is performed for a single event (E_j) at a time.

Where:

The kernel function for observation comparisons, has two levels. The first level, indicated by the values, differentiates potentially mappable observation pairs from non-mappable observation pairs. The second level takes into account the temporal congruence of the system and reference event observations and the observation’s detection score in relation to the system’s range of detection score. The decision scores are taken into account to facilitate the DET curve generation. By giving more weight to higher confidence score observations, realignment can be avoided during DET curve production.

4.3. Detection Error Tradeoff Curves

Graphical performance assessment uses a Detection Error Tradeoff (DET) curve that plots a series of event-averaged missed detection probabilities and false alarm rates that are a function of a detection threshold, Θ. This Θ is applied to the system’s detection scores meaning the system observations with scores above the Θ are ‘declared‘ to be the set of detected observations. After Θ is applied, the measurements are then computed separately for each event, then averaged to generate a DET line trace. The per-event formulas for P_Miss and R_FA are:

Where

The formulas to compute averages over all events are defined as:

PMiss(S, Θ) is not defined for all events because N_Targ(E_i) may be 0. Therefore PMiss(S, Θ) is calculated over the set of events with true occurrences. This enables the evaluation of a system on events that do not exist in the test corpus.

4.4. DCR Computations

The evaluation will use the Normalized Detection Cost Rate (NDCR) measure for evaluating system performance. NDCR is a weighted linear combination of the system’s Missed Detection Probability and False Alarm Rate (measured per unit time). The measure’s derivation can be found in Appendix A and the final formula is summarized below. NIST will report an NDCR for each event and not average them over events.

Where:

The measure’s unit is in terms of Cost per Unit Time which has been normalized so that an NDCR=0 indicates perfect performance and an NDCR=1 is the cost of a system that provides no output, i.e. P_Miss=1 and P_FA=0.

For the 2009 evaluation, Cost_FA = 1, Cost_Miss = 10, and R_Target = 20.0.

Two versions of the NDCR will be calculated for each system: the Actual NDCR and the Minimum NDCR.

4.4.1. Actual NDCR

The Actual NDCR is the primary evaluation metric. It is computed by restricting the putative observations to those with true actual decisions.

4.4.2. Minimum NDCR

The Minimum NDCR is a diagnostic metric. It is found by searching the DET curve for the Θ with the minimum cost. The difference between the value of Minimum NDCR and Actual NDCR indicates the benefit a system could have gained by selecting a better threshold.

5. Events

Initially, a video event is defined to be “an observable action or change of state in a video stream that would be important for airport security management”. Events may vary greatly in duration, from 2 frames to longer duration events that can exceed the bounds of the excerpt.

Events will be described through an “event description document”. The document will include a textual description of the event and a set of exemplar event occurrences (annotations). Each exemplar will indicate the source file and temporal coordinates of the event. Events will be considered to be independent for the evaluation. Therefore, systems may build separately trained models for each event.

There will be two sets of events: Required events and Optional events. There is no implicit difference between the types of events included in the event sets. Systems should output detection results for any three events in the required set. For the optional set, systems can output detection results for some or all of the events.

The required event set is as follows:

PersonRuns, CellToEar, ObjectPut, PeopleMeet, PeopleSplitUp, Embrace, Pointing, ElevatorNoEntry, OpposingFlow, and TakePicture (defined in the “event description document”).

6. Submission of results

Submissions will be made via ftp according to the instructions in Appendix B. In addition to the system output, a system description is also required for each condition. This description must include a description of the hardware used to process the data, and a detailed description of the architecture and algorithms used in the system.

7. Schedule

The 2009 evaluation schedule for event detection includes the following milestones:

Jan.--Mar.: Event detection planning & telecons

Feb.: Call for participation in TRECVid

Mar.: Development and Evaluation data specified

Mar.: Release of video data, required event definitions, and examples

Mar.: Final evaluation plan & guidelines written

Mar.: Release scoring tool

Mar.: Development annotations for required events released

Jul.: Dry run (systems run on Dev data)

Sep.: Obtain submissions for formal evaluation

Oct. 1: Distribute preliminary results

Oct. 10: Distribute final results

Oct.: Notebook papers due at NIST

Nov.: Present results at TRECVid

8. References

[1] Harold W. Kuhn, "The Hungarian Method for the assignment problem", Naval Research Logistic Quarterly, 2:83-97, 1955.

[2] Martin, A., Doddington, G., Kamm, T., Ordowski, M., Przybocki, M., “The DET Curve in Assessment of Detection Task Performance”, Eurospeech 1997, pp 1895-1898.

Appendix A: Derivation of Average Normalized Detection Cost Rate

Average Normalized Detection Cost Rate (ANDCR) is a weighted linear combination of the system’s Missed Detection Probability and False Alarm Rate (measured per unit time). The constant parameters of ANDCR, which are specified below, represent both the richness of events in the source data and the relative detriment of particular error types to a hypothetical application.

The cost of a system begins with the cost of missing an event (Cost_Miss) and the cost of falsely detecting an event (Cost_FA). N_Miss(S,E) is the number of missed detections for system S, event E. N_FA(S,E) is the number of false alarms for the same system and event.

To facilitate comparisons across systems and test sets, we convert Detection Cost to a rate by dividing by the length of the source data. Typically, we make this conversions to percentages by dividing by the count of discrete units for which systems make decisions. In a streaming environment, there are no discrete units, therefore normalizing by unit time is a more appropriate normalization. Note also that the measure of Type I error, R_FA, is commonly used in surveillance-style applications.

R_Target(E) is the rate of occurrences for the event. This value is dependent on the event but providing this prior to a system for each event changes the definition of an event – it includes the event definition and the prior. Instead, we replace the event-dependent prior with a single, global prior, R_Target, that in combination with the Cost_Miss and Cost_FA reflects the characteristics of the surrogate application. While the events for the evaluation have not been selected yet, we expect the them to have similar numbers of occurrences[1]: neither too frequent or too rare. Therefore, the single prior is warranted. The modified formula becomes:

The range of the DCR_Sys measure is [0,∞). To ground the costs, a second normalization scales the cost to be 0 for perfect performance and 1 to be the cost of a system that provides no output (therefore P_Miss = 1 and P_FA = 0). The resulting formula is the Normalized Detection Cost Rate of a system (NDCR).

Where:

Beta is separated out because it is composed of constant values that define the parameters of the surrogate application.

To calculate performance over an ensemble of events, we define Average Normalized Detection Cost Rate (ANDCR(S)) by averaging the Missed Detection probabilities for all events with at least one true event occurrence (N_EventsNZ) and averaging the False Alarm Rates overall events. By separating the two averages, the measure can incorporate events with no true occurrences while remaining defined.

Where:

The measure’s unit is in terms of Cost per Unit Time.

Appendix B: Submission Instructions

The packaging and file naming conventions for the TRECVid 2009 Event Detection evaluation relies on Experiment Identifiers (EXP-ID) to organize and identify the files for each evaluation condition and link the system inputs to system outputs. Since EXP-IDs may be used in multiple contexts, some fields contain default values. The following section describes the EXP-IDs to be used for the London Gatwick airport surveillance development dataset (devset) and evaluation dataset (evalset).

The following BNF describes the EXP-ID structure:

EXP-ID ::= <SITE>_<YEAR>_<TASK>_<DATA>_<LANG>_<INPUT>_

where,

<SITE> ::= expt | short name of participant’s site

The special SITE code “expt” is used in the EXP-ID to indicate a reference annotation.

<YEAR> ::= 2009

<TASK> ::= retroED

<DATA> ::= DEV09 | EVAL09

<LANG> ::= ENG

<INPUT> ::= s-camera

<SYSID> ::= a site-specified string (that does not contain underscores) designating the system used

The SYSID string must be present. It is to begin with p- for a primary system or with c- for any contrastive systems. For example, this string could be p-baseline or c-contrast. This field is intended to differentiate between runs for the same evaluation condition. Therefore, a different SYSID should be created for runs where any changes were made to a system.

<VERSION> ::= 1..n (with values greater than 1 indicating multiple runs of the same experiment/system)

In order to facilitate transmission to NIST and subsequent scoring, submissions must be made using the following protocol, consisting of three steps: (1) preparing a system description, (2) packaging system outputs and system descriptions, and (3) transmitting the data to NIST.

B.1 System Descriptions

Documenting each system is vital to interpreting evaluation results. As such, each submitted system, (determined by unique experiment identifiers), must be accompanied by a system description with the following information:

Section 1. Experiment Identifier(s)

List all the experiment IDs for which system outputs were submitted. Experiment IDs are described in further detail below.

Section 2. System Description

A brief technical description of your system; if a contrastive test, contrast with the primary system description.

List all events processed on a single line as follows:

Events_Processed: Event1 Event2 Event3 …

(Note: each event should come from the required event set.)

Section 3. Training:

A list of resources used for training and development.

Section 4. References:

A list of all pertinent references.

B.2 Packaging Submissions

All system output submissions must be formatted according to the following directory structure:

output/<EXP-ID>/<EXP-ID>.txt

output/<EXP-ID>/<SOURCE_FILE>*.xml

where,

EXP-ID is the experiment identifier,

<EXP-ID>.txt is the system description file as specified above (section B.1),

< SOURCE_FILE >*.xml is the set of ViPER-formatted system output files generated for the indicated experiment. The SOURCE_FILE stem corresponds to the name of the video, without its extension. The events detected for each source file must be combined into a single ViPER file. Thus, for a given EXP-ID, there will be an experimental control file (ECF), and several system output files, one for each video referenced in the ECF. (Note: the evaluation tools will contain a ViPER file merging script ‘TV08MergeHelper’, useful for combining ViPER files that may contain only observations for a single event into a larger ViPER file.)

For example, the ECF ‘expt_2009_retroED_DEV08_ENG_s-camera_NIST_1.ecf.xml’ will have the following system outputs (from site SITE, system SYS), in a directory named SITE_2009_retroED_DEV09_ENG_ s-camera_NIST_1:

SITE_2009_retroED_DEV09_ENG_s-camera_SYS_1.txt
LGW_20071123_E1_CAM1.xml
LGW_20071123_E1_CAM2.xml
LGW_20071123_E1_CAM3.xml
LGW_20071123_E1_CAM4.xml
LGW_20071123_E1_CAM5.xml
…

B.3 Transmitting Submissions

To prepare your submission, first create the previously described file/directory structure. This structure may contain the output of multiple experiments, although you are free to submit one experiment at a time if you prefer. The following instructions assume that you are using the UNIX operating system. If you do not have access to UNIX utilities or ftp, please contact NIST to make alternate arrangements.

First, change directory to the parent directory of your “output/” directory. Next, type the following command:

tar -cvf - ./output | gzip > <SITE>_<SUB-NUM>.tgz

where,

<SITE> is the ID for your site

<SUB-NUM> is an integer 1 to n, where 1 identifies your first submission, 2 your second, etc.

This command creates a single tar/gzip file containing all of your results. Next, ftp to jaguar.ncsl.nist.gov giving the username 'anonymous' and (if requested) your e-mail address as the password. After you are logged in, issue the following set of commands, (the prompt will be 'ftp>'):

ftp> cd incoming

ftp> binary

ftp> put <SITE>_<SUB-NUM>.tgz

ftp> quit

Note that because the “incoming” ftp directory (where you just ftp’d your submission) is write protected, you will not be able to overwrite any existing file by the same name (you will get an error message if you try), and you will not be able to list the incoming directory (i.e., with the “ls” or “dir” commands). Please note whether you get any error messages from the ftp process when you execute the ftp commands stated above and report them to NIST.

The last thing you need to do is send an e-mail message to jfiscus@nist.gov, travis.rose@nist.gov, and martial@nist.gov to notify NIST of your submission. The following information should be included in your email:

the name of your submission file,
the file size,
a listing of each of your submitted experiment IDs.

Please submit your files in time for us to deal with any transmission errors that might occur well before the due date if possible. Submissions must validate against TrecVid08.xsd, or they will be rejected. Note that submissions received after the stated due dates for any reason will be marked late.

B.4 Scoring Submissions

To score system output against ground truth, the TV08Scorer (part of F4DE software) is invoked as follows:

TV08Scorer --showAT --allAT --writexml <writexml_dir> --pruneEvents

--computeDETcurve --OutputFileRoot <outputfileroot_filebase>

--titleOfSys "<titleofsys_string>" --observationCont --ecf <ecf_file>

<submissionsysfiles_list> --gtf <referencefiles_list> --fps 25 --deltat 10

--limitto <event1,event2,event3> --MissCost 10 --CostFA 1 --Rtarget 20

where:

- <writexml_dir> is a directory in which alignment ViPER files will be written

(containing Mapped, Unmapped_Sys, Unmapped_Ref Events) (example value: alignment_xmls)

- <outputfileroot_filebase> is the file base for all DCR output (example value: Dcr_results)

- <titleofsys_string> is the string that will be added to the plots (example value:

<SITE> | <SYS-ID>)

- <ecf_file> is the ECF (XML) file against which the scoring is made (note: all the files listed within the sys or ref files have to be listed in the ECF for scoring to occur) (example value: expt_2008_retroED_DEV08_ENG_s-camera_p-baseline_1.ecf)

- <submissionsysfiles_list> is the list of SYS files provided (example from dryrun:

sys/LGW_20071112_E1_CAM1.xml sys/LGW_20071112_E1_CAM2.xml

sys/LGW_20071112_E1_CAM3.xml sys/LGW_20071112_E1_CAM4.xml

sys/LGW_20071112_E1_CAM5.xml)

- <referencefiles_list> is the list of REF files to score against (example from dryrun: ref/LGW_20071112_E1_CAM1.xgtf ref/LGW_20071112_E1_CAM2.xgtf

ref/LGW_20071112_E1_CAM3.xgtf ref/LGW_20071112_E1_CAM4.xgtf

ref/LGW_20071112_E1_CAM5.xgtf)

- <event1,event2,event3> is the list of events scored here and should be the same list as found in the submission's "Events_Processed:" line but comma separated here instead of space separated.

Appendix C: Experimental Control File (ECF)

Experiment Control Files (ECF)s are the mechanism the evaluation infrastructure uses to specify the source file excerpts to use for a specific experimental condition. The ECF file specifies a several attributes of excerpt such as the time regions within a media file, the language, and the source type.

Two ECF files will be used for the evaluation. The first is a system input ECF file that will be provided for all tasks to indicate what data is to be processed by the system. The second is a scoring ECF file that the evaluation code uses to determine the range of data to evaluate the system on. In the event a problem is discovered with the data, a special scoring ECF file will be used to specify the time regions to be scored or excluded from scoring.

1. ECF File Naming

ECF file names use the relevant EXP-ID and end with the ‘.ecf.xml’ extension.

2. ECF File Format Description

An ECF file is an XML-formatted file that consists of a hierarchy of nodes, “ecf”, “excerpt_list”, and “excerpt”, described below:

The “ecf” node contains an “excerpt_list” node and the following elements:

· source_signal_duration: a floating point number indicating the total duration in seconds of the video recording specified by the excerpts

· version: a version identifier for the ECF file

· excerpt_list: node contains a set of “excerpt” nodes.

The “excerpt” element is a non-overlapping node that specifies the excerpt from a media file to be used in the evaluation. The “excerpt” node has no attributes. The “excerpt” node has the following elements:

· filename: this attribute indicates the source file.

· language: the language of the source file (“english”).

· source_type: the source type of the recording (“surveillance”).

· begin: the beginning time of the segment to process. The time is measured in seconds from the beginning of the recording which is time = 0.0

· duration: the duration of the excerpt measured in seconds.

3. An example ECF XML fragment:

<ecf>

<source_signal_duration>7200.00</ source_signal_duration >

<excerpt_list>

<language>english</language>

<source_type>surveillance</source_type>

</excerpt>

<language>english</language>

<source_type>surveillance</source_type>

</excerpt>

…

<language>english</language>

<source_type>surveillance</source_type>

</excerpt>

…

</excerpt_list>

…

</ecf>

[1] For instance within 2-3 orders of magnitude.