Ad-hoc Video Search (AVS)

Task Coordinators: Georges Quénot and George Awad

The Ad-hoc search task goal is to model the end user search use-case, who is searching (using textual sentence queries) for segments of video containing persons, objects, activities, locations, etc. and combinations of the former. While the Internet Archive (IACC.3) dataset was adopted between 2016 to 2018, from 2019 to 2021 a new data collection (V3C1) based on Vimeo Creative Commons (V3C) datset was adopted. Starting in 2022 the task started to utilize a new sub-collection V3C2 to test systems on a new set of queries in addition to common (fixed) progress query set to measure system progress from 2022 to 2024.

System Task

Given the video test collection (V3C2), master shot boundary reference, and set of textual Ad-hoc queries (approx. 30 queries) released by NIST, return for each query a list of at most 1000 shot IDs from the test collection ranked according to their likelihood of containing the target query.

Two evaluation tasks will be supported: main task with new queries annually, and a progress subtask to measure system progress across 2 and 3 years respectively.

1- Main Task (2022 - 2024):

Systems will be asked to return results for *New unseen* queries annually as follows:
30 New queries in 2022, where all 30 queries will be evaluated and scored
20 New queries in 2023, where all 20 queries will be evaluated and scored
20 New queries in 2024, where all 20 queries will be evaluated and scored

2- Progress Subtask (2022 - 2024):

Systems will be asked to return results for 20 Common (fixed) queries annually from 2022 to 2024 where evaluation schedule will be as follows:

Although systems are asked to submit results in 2022 for the 20 common queries, no evaluation for them will be conducted in 2022. NIST will just save the runs for subsequent years.
NIST will evaluate and score, in 2023, subset of 10 out of the 20 common queries submitted in 2022 AND 2023 to allow comparison of performance between the two years.
NIST will evaluate and score, in 2024, the other 10 subset out of the 20 common queries submitted in 2022, 2023 AND 2024 to allow comparison of performance across the three years.

Given the above schedule of query distribution from 2022 to 2024, systems are encouraged to join both tasks and plan to submit results for:
50 (30 New + 20 common) queries in 2022
40 (20 New + 20 common) queries in 2023 (current year)
40 (20 New + 20 common) queries in 2024

Data Resources

The testing dataset
Vimeo Creative Commons Collection (V3C2) is 9760 videos (1.6 TB, 1300 total hours) with mean video duration of 8 min and total 1,425,454 video segments.
The master shot reference and an associated readme file are available for download.

The development dataset

2016-2018 Internet Archive (IACC.3) dataset (Used by the Ad-hoc Video Search (AVS) task) of 4593 Internet Archive videos (144GB, 600 total hours) using videos with durations between 6.5min and 9.5min. See the information provided here for download instructions and past ground truth.
2013-2015 (Internet Archive IACC.2.A, IACC.2.B, and IACC.2.C datasets) used by the Semantic Indexing task each containing about 200 h drawn from the IACC.2 collection using videos with durations ranging from 10s to 6.4 min. These datasets can be downloaded; see the information provided here
2010-2012 (Internet Archive IACC.1.tv10.training, IACC.1.A, IACC.1.B, and IACC.1.C datasets) used by the Semantic Indexing task 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; see the information provided here
Examples of previous years' Ad-hoc queries are available: (No system training is allowed using the V3C2 testing dataset)
2016 Queries
2017 Queries
2018 Queries
2019 Queries
2020 Queries
2021 Queries
2022 Queries
2023 Queries
All past years' ground truth data are available from our past data webpage

Previous collaborative annotations:
The results of past collaborative annotations on Sound and Vision as well as Internet Archive videos from 2007-2013 are available for use in system development.
Sharing of components:

Previous and current teams are encouraged to share system components (concept scores, model outputs, codebase resources, etc) with other teams to facilitate new team's experience and get better insights into system behaviour through broader experiments.
- The Centre for Research and Technology Hellas (ITI-CERTH) team shared their concept detection scores for the IACC.3 dataset
- Frame-level CNN features for IACC.3 dataset are available by RUCMM team here
- A set of classification results, features and high-level analysis of the V3C1 dataset were generated and provided by Klagenfurt University and University of Basel. Similar results and analysis for V3C2 dataset is available from HERE
- RUCMM team has made available the V3C1-PC (V3C1-Pseudo-Caption) pre-training auto-generated dataset which consists of 436,204 captions for 219,531 video shots in the V3C1 collection.

Participation types

There will be 3 types of participation:

Participation by submitting only automatic runs for TRECVID evaluation
Participation by submitting only automatic runs for TRECVID evaluation and by using an interactive system during the next VBS (Video Browser Showdown)
Participation by using only interactive system during the next VBS (Video Browser Showdown)

For questions about participation in the next VBS please contact the VBS organizers: Werner Bailer, Cathal Gurrin, or Klaus Schoeffmann.

Allowed training categories:

Four training types are allowed in general. Each run must declare which training type it used to generate it's results.

A - used only V3C1 data and annotation strict.
D - used any other training data with any annotation (except the testing data V3C2).
E - used only training data collected automatically using only the official query textual description.
F - used only training data collected automatically using a query built manually from the given official query textual description.

The training categories E and F support experiments using a no annotation condition. The idea is to promote the development of methods that permit the indexing of concepts in video shots using only data from the Web or archives without the need of additional annotations. The training data could for instance consist of images or videos retrieved by a general purpose search engine (e.g. Google) using only the query definition with only automatic processing of the returned results.

By "no annotation", we mean here that no annotation should be manually done on the retrieved samples (either images or videos). Any annotation done by somebody else prior to the general search does not count. Methods developed in this context could be used for building indexing tools for any concept starting only from a simple query defined for it.

Please note these restrictions and information on training types.

Run submission types:

Three main submission types will be accepted:

Fully automatic (F) runs (no human input in the loop): A System takes official query as input and generates results without any human intervention.
Manually-assisted (M) runs: where a human can formulate the initial query once based on topic and query interface, not on knowledge of collection or search results. Then the system takes the formulated query as input and generates results without further human intervention.
Relevance-Feedback (R) runs: System takes the official query as input and generates initial results, then a human judge can assess up to the top-30 results and input this information as a feedback to the system to generate a new set of results. This feedback loop is strictly permitted ONLY up to 3 iterations.

An extra 1 run of a novelty type (N) is allowed to be submitted within the main task. The goal of this run is to encourage systems to submit novel and unique relevant shots not easily discovered by other runs. Each team may submit only 1 novelty run. Please note the required xml field in the dtd file indicating if the run is of novelty or common type.

Optional attributes are available for teams to submit explainability results. This is supported in the xml run files as optional timestamp and bounding box coordinates (please see the XML DTD container for run files) to localize "why" the submitted shot is relevant to the query. The goal of this initiative is to allow for more detailed diagnostic results for teams and NIST and to assess if submitted results are relevant for the correct reasons.

Given the two main current paradigms in video search (concept-based vs visual-semantic embedding), a new run type will be supported for submission to better learn the differences, strength and weaknesses of these two main approaches. Systems can submit an additional 1 run ONLY based on concept detection per submission type, thus maximum of 3 concept-based runs (if submitting using the 3 submission types F, M and R). Please consult the DTD file for labeling the concept-based run in your xml submission file

The submission types (automatic, manually-assisted, relevance feedback) are orthogonal to the training types (A, D, E, F). Each team may submit a maximum of 4 prioritized runs, per submission type and per task type (Main or Progress), with 2 additional if they are of the "no annotation" training type (E or F) and the others are not. 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 and correct if needed 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.Counter -v YourSubmision.xml.

We highly encourage each team to submit 1 baseline run for their system approach so comparison and analysis with other advanced runs can be conducted. Please when submitting the baseline run, make sure to label the run file name with a clear label (eg. "baseline").

Run submission format:

Participants will submit results against V3C2 data in each run for all and only the 20 main queries or the 20 progress queries released by NIST and for each query at most 1000 shot IDs.

The DTD includes an xml field to determine the task type (Main or Progress) as well as the run types. Each of the Main and Progress tasks will have their own set of Query IDs to use within run submissions.

Here for download (right click and choose "display page source" to see the entire files) is a DTD file for Adhoc search results of one main run, the container for one run, and a small example of what a site would send to NIST for evaluation. Please check all your submissions to see that they are well-formed.

Please submit each of your runs in a separate file, named to make clear which team has produced it. EACH file you submit should begin, as in the example submission, with the DOCTYPE statement:

that refers to the DTD at NIST via a URL and with a videoAdhocSearchResults element even though there is only one run included.

Remember to use the correct shot IDs in your submissions. The shot IDs take the form of "shotXXXX_YY", where XXXX is the original video ID and YY is the segmented shot ID. Please don't use any keyframe associated file names in your submissions. Consult the V3C2 readme file for more information about submitting the correct shot file names and the master shot reference for V3C2 dataset

Please note the *optional* explainability attributes with each submitted shot to indicate a timestamp (in seconds) of a frame within the shot and a bounding box coordinates.

All runs will be submitted using password-protected webpage prepared by NIST to upload your runs.

Please consult the general TRECVID schedule page for query release and run submission due dates

Evaluation:

All 2024 main 20 queries will be evaluated by assessors at NIST after pooling and sampling.
For the progress subtask, NIST will evaluate and score 10 queries to compare progress from 2022 to 2024.
Please note that NIST uses a number of rules in manual assessment of system output.

Measures:

Mean extended inferred average precision (mean xinfAP), will be applied. This allows sampling density to vary e.g. so that it can be 100% in the top strata, which are most important for average precision.
As in past years, other detailed measures based on recall, precision will be provided by the sample_eval software.
Speed will also be measured: clock time per query search, reported in seconds (to one decimal place) must be provided in each run.
A special metric will be applied to score Novelty runs such that more credit can be given to unique shots.