ActEV: Activities in Extended Video


Datasets

Framework

The DIVA Framework is a software framework designed to provide an architecture and a set of software modules which will facilitate the development of activity recognition analytics. The Framework is developed as a fully open source project on GitHub. The following links will help you get started with the framework: The DIVA Framework is based on KWIVER, an open source framework designed for building complex computer vision systems. The following links will help you learn more about KWIVER:
  • KWIVER Github Repository This is the main KWIVER site, all development of the framework happens here.
  • KWIVER Issue Tracker Submit any bug reports or feature requests for the KWIVER here. If there's any question about whether your issues belongs in the KWIVER or DIVA framework issues tracker, submit to the DIVA tracker and we'll sort it out..
  • KWIVER Main Documentation PageThe source for the KWIVER documentation is maintained in the Github repository using Sphinx. A built version is maintained on ReadTheDocs at this link. A good place to get started in the documentation, after reading the Introduction are the Arrows and Sprokit sections, both of which are used by the KWIVER framework.

Baseline Algorithms

KITWARE has adapted two "baseline" activity recognition algorithms to work within the DIVA Framework:

Visualization Tools

Annotation Tools

Contact Us

For information on data, evaluation code, etc., please email: actev-nist@nist.gov

For ActEV evaluation discussion, please visit our Google Group: https://groups.google.com/a/list.nist.gov/forum/#!forum/trecvid.actev

Activity Examples

An ActEV activity is defined to be “one or more people performing a specified movement or interacting with an object or group of objects”. Activities are annotated by humans using a set of annotation guidelines that specify how to perform the annotation and the criteria to determine if the activity occurred. Each activity is formally defined by five elements:
  • Activity Name - A mnemonic handle for the activity
  • Activity Description - Textual description of the activity
  • Begin time rule definition - The specification of what determines the beginning time of the activity
  • End time rule definition - The specification of what determines the ending time of the activity
  • Required object type list - The list of objects systems are expected to identify for the activity. Note: this aspect of an activity not addressed by ActEV-PC.

For example:

Activity Name
Description and Example Chip Videos

person_closes_vehicle_door
  • Description: A person closing the door to a vehicle.
  • Start: The event begins 1 s before the door starts to move.
  • End: The event ends after the door stops moving. People in cars who close the car door from within is a closing event if you can still see the person within the car. If the person is not visible once they are in the car, then the closing should not be annotated as an event.
  • Objects associated with the activity : Person; and Door or Vehicle


vehicle_turns_left
  • Description: A vehicle turning left or right is determined from the POV of the driver of the vehicle. The vehicle may not stop for more than 10 s during the turn.
  • Start: Annotation begins 1 s before vehicle has noticeably changed direction.
  • End: Annotation ends 1 s after the vehicle is no longer changing direction and linear motion has resumed. Note: This event is determined after a reasonable interpretation of the video.
  • Objects associated with the activity : Vehicle


person_loads_vehicle
  • Description: An object moving from person to vehicle.
  • Start: The event begins 2 s before the cargo to be loaded is extended toward the vehicle (i.e., before a person’s posture changes from one of “carrying” to one of “loading”).
  • End: The event ends after the cargo is placed into the vehicle and the person-cargo contact is lost. In the event of occlusion, it ends when the loss of contact is visible.
  • Objects associated with the activity: Person; and Vehicle


The names of the 37 Known Activities for ActEV’21 SDL


person_abandons_package person_loads_vehicle person_stands_up
person_closes_facility_door person_transfers_object person_talks_on_phone
person_closes_trunk person_opens_facility_door person_texts_on_phone
person_closes_vehicle_door person_opens_trunk person_steals_object
person_embraces_person person_opens_vehicle_door person_unloads_vehicle
person_enters_scene_through_structure person_talks_to_person vehicle_drops_off_person
person_enters_vehicle person_picks_up_object vehicle_picks_up_person
person_exits_scene_through_structure person_purchases vehicle_reverses
person_exits_vehicle person_reads_document vehicle_starts
hand_interacts_with_person person_rides_bicycle vehicle_stops
person_carries_heavy_object person_puts_down_object vehicle_turns_left
person_interacts_with_laptop person_sits_down vehicle_turns_right
vehicle_makes_u_turn
Updates
  • Sep 21, 2020: The ActEV'21 SDL UF with Known Activities opens
  • Sep 21, 2020: ActEV SDL UF is a competition under the WACV HADCV'21 workshop
Summary
ActEV is a series of evaluations to accelerate the 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. Currently we are running the ActEV 2021 Sequestered Data Leaderboard (SDL) evaluation that features Unknown Facility and Surprise Activity Testing and the ActEV TRECVID 2020 evaluation that features additional known activities for a known facility.
Past Evaluations
ActEV began with the Summer 2018 Blind and Leaderboard evaluations for 12 activities. The summer evaluation was followed by the ongoing Fall ActEV Self-Reported Evaluation which ended in Dec 2018 and included 18 activities. The Activities in Extended Videos Prize Challenge (ActEV-PC) ran under CVPR'19 ActivityNet workshop. In 2019 we ran two other evaluations, the ActEV 2019 Sequestered Data Leaderboard (SDL) and the ActEV TRECVID 2020 evaluations.
What is Activity Detection in Extended Videos?
An ActEV activity is defined to be “one or more people performing a specified movement or interacting with an object or group of objects”. Activity detection technologies process extended video streams, such as those from a CCTV camera, and automatically detects all instances of the activity by: (1) identifying the type of activity, (2) producing a confidence score indicating the presence of instance, (3) temporally localizing the instance by indicating the begin and end times, and (4) optionally, detecting and tracking the objects (people, vehicles, objects) involved in the activity.
Click on the tabs above to see video examples, activity examples, and evaluation tasks
What
The ActEV evaluations are being conducted to assess the robustness of automatic activity detection for a multi-camera streaming video environment.
Who
Everyone. Anyone who registers can submit to the evaluation server.
How
Register here and then based on the evaluation participants can either ran their activity detection software on their compute hardware and submit their system output to the ActEV Scoring Server or submited their runnable activity detection software to NIST using the Evaluation Commandline Interface. See the individual evaluation pages and evaluation plans for details.
Data
Each ActEV evaluation uses a new video data set, changes the evaluation tasks, or adds/changes activities. The data will be provided in MPEG-4 and AVI formatted files. See the individual evaluation pages for details.
Evaluation Metrics and Tools
The main scoring metrics will be based on detection, temporal localization, and spatio-temporal localization using evaluation measures that include the probability of missed detection and rate of false alarm. See details in the evaluation plans of each evaluation.

NIST maintains the ActEV Scoring Software on the Scoring software for the Activities in Extended Video (ActEV) evaluation GitHub repo.

ActEV: Video Examples
Below you will find four example videos from our data sets. There are two example views each of indoor and outdoor.
Location View 1 View 2
Indoor
Outdoor
ActEV Evaluation Tasks
Activity detection has been researched for many years and remains an unsolved computer vision challenge that requires many capabilities beyond the current state of the art. The ActEV series supports several evaluation tasks each escalating the difficulty by requiring more specific information from the system. Presently, there are three evaluation tasks defined: 1) Activity Detection (AD), 2) Activity and Object Detection (AOD), and (3) Activity and Object Detection and Tracking (AODT). Each evaluation task is summarized below. For a full description of the evaluation tasks, read the Evaluation Plan for each specific evaluation.
Activity Detection (AD)

For the Activity Detection task, given a target activity, a system automatically detects and temporally localizes all instances of the activity. For a system-identified activity instance to be evaluated as correct, the type of activity must be correct and the temporal overlap must fall within a minimal requirement.
Activity and Object Detection (AOD)

For the Activity and Object Detection task, given a target activity, a system detects and temporally localizes all instances of the activity and spatially detects/localizes the people and/or objects associated with the target activity. For a system-identified instance to be scored as correct, it must meet the temporal overlap criteria for the AD task and in addition meet the spatial overlap of the identified objects during the activity instance.
Activity Object Detection and Tracking (AODT)

For the Activity Object Detection and Tracking task, given a target activity, a system detects and temporally localizes all instances of the activity, spatio-temporally detects/localizes the people and/or objects associated with the target activity, and properly assigns IDs the objects play in the activity. For a system-identified instance to be scored as correct, it must meet the temporal overlap criteria and spatio-temporal overlap of the objects for the AOD task and correctly assign the IDs to the objects as described in the activity definition.