The Eyelashes dataset is the first publicly available dataset of 3D eyelashes containing a total of 56 subjects.

The dataset is intended to be used for evaluation, validation, or training models to predict eyelashes information on images, meshes, texts or videos. A detailed description of the dataset can be found on our 

 page. The license conditions are mentioned on the 

Documents and papers that report on research that uses the Eyelashes Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:

Glenn Kerbiriou, Quentin Avril, Marchal, Maud. « 3D Reconstruction and Semantic Modeling of Eyelashes” Eurographics (2024)

It would be highly appreciated if this use was shared with Interdigital, at this address: 

, with “Eyelashes” in the subject of your email.

Eyelashes Dataset

The DisplayPower2k dataset is a set of power consumption measures for different OLED screens, and pixel-wise power map images, computed from 648 images.

Additional energy-aware images and their corresponding power measures are also distributed together with the DisplayPower2k dataset.

The dataset is intended to be used for training and assessing models to predict and control the power consumption of images when displayed on screens. A detailed description of the benchmark can be found on our 

. The license conditions are mentioned on the 

All documents and papers that report on research that uses the DisplayPower2k Dataset must acknowledge the use of the dataset by including an appropriate citation to the followings:

Claire-Hélène Demarty, Laurent Blondé, Olivier Le Meur. "Display Power Modeling for Energy Consumption Control." ICIP (2023).

Olivier Le Meur, Claire-Hélène Demarty, Laurent Blondé. "Deep-Learning-based Energy Aware Images." ICIP (2023).

It would be highly appreciated if this use was shared with InterDigital, at this address: 

, with “DisplayPower2k” in the subject of your 

DisplayPower2k dataset

The FilmGrainStyle740k dataset is a set of image pairs of grain-free and grainy images. Images with grain were generated from the clean images, using some proprietary plug-in in the VFX and Film Editing Software Nuke, that allows to adjust the grain patterns to the underlying content according to three attributes from the original images: intensity, texture and saturation. This plugin was elaborated with support from professional colorists. For each grain-free image, five grainy versions were generated.

The dataset is intended to be used for training deep learning models for film grain analysis, film grain removal and synthesis, film grain detection, film grain intensity prediction, film grain quality assessment, etc. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

This Dataset comes from a collaboration with the IETR Lab (

All documents and papers that report on research that uses the FilmGrainStyle740k Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:

Zoubida Ameur, Claire-Hélène Demarty, Olivier Le Meur, Daniel Ménard and Edouard François. "Style-based film grain analysis and synthesis." ACM Multimedia Systems (2023).

filmgrainstylemanagement@interdigital.com

, with “FilmGrainStyle740k” in the subject of your email.

FilmGrainStyle740k Dataset

The FilmGrain Dataset is a set of image pairs with and without film grain (grainy / grain-free) at five different film grain intensity levels.

It is intended to be used for training a deep learning model for film grain removal and synthesis, film grain detection, film grain intensity prediction, etc. A detailed description of the benchmark can be found on our Data 

All documents and papers that report on research that uses the FilmGrain Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:

Z. Ameur, W. Hamidouche, E. François, M. Radosavljević, D. Menard, C.-H. Demarty. Deep-based Film Grain Removal and Synthesis. 

Title = { Deep-based Film Grain Removal and Synthesis },

Author ={Zoubida Ameur and Wassim Hamidouche and Edouard François and Milo\v{s} Radosavljevi\’{c} and Daniel M\’{e}nard and Claire-H\'{e}l\`{e}ne Demarty},

, with “FilmGrain” in the subject of your email.

FilmGrain Dataset

10 sequences were created using the Unity game engine to load and play several games. We captured the frame buffers using Unity Recorder and custom post-processing shaders in both the built-in render pipeline and the High-Definition Render Pipeline. We captured the color buffer into an RGBA16 Signed Float texture as well as the depth and optical flow in a second similar texture (with the linear depth info encoded into the R channel, and the [X,Y] coordinates of the motion vectors into the [G,B] channels). The textures were recorded at 60FPS in 1080p and 4K resolutions and written into uncompressed EXR 16 bit images. Two YUVs sequences were then generated. The depth is stored in the Y channel and converted to 14 bits to avoid high bit-depth coding, vertical and horizontal motion vectors are stored at quarter pixel precision in U/V channels for the corresponding texture frame.

Figure 1 /3D_game _kit_Level-1 sequence Frame #10 top left YUV texture, top right, depth information (Y), bottom left vertical motion (V), bottom right horizontal motion (U)

Gaming Sequences

The Long Range (LoRa) protocol for low-power wide-area networks (LPWANs) is a strong candidate to enable the massive roll-out of the Internet of Things (IoT) because of its low cost, impressive sensitivity (-137dBm), and massive scalability potential. As tens of thousands of tiny LoRa devices are deployed over large geographic areas, a key component to the success of LoRa will be the development of reliable and robust authentication mechanisms. We publicly share waveform data from 100 bit-similar devices (with identical manufacturing processes) over different deployment scenarios (outdoor vs. indoor) and spanning several days.

LoRa Radio Data

In contrast to existing datasets with very few video resources and limited accessibility due to copyright constraints, LIRIS-ACCEDE consists of videos with a large content diversity annotated along affective dimensions. All excerpts are shared under Creative Commons licenses and can thus be freely distributed without copyright issues. The dataset (the video clips, annotations, features and protocols) is publicly available.

LIRIS-ACCEDE

LaFin: Large-scale Flickr interestingness

 dataset (hereafter “the Dataset”) is a collection of Flickr image IDs corresponding to about 123k Flickr images, equally balanced between interesting and non-interesting images, and their corresponding metadata. In addition to the images, their binary labels, and associated metadata, some precomputed features are provided: CNNs, semantic features that derived from image captioning and Word2Vec representations of Flickr tags. 

It is intended to be used for analyzing socially-driven image interestingness and building a machine learning model for prediction. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

All documents and papers that report on research that uses the LAFIN Image Interestingness Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:

E.Berson, N. Q. K. Duong and C.-H. Demarty. Collecting, Analyzing and Predicting Socially-Driven Image Interestingness. In Proceedings of the EUSIPCO Conference, Spain, 2019.

Title = { Collecting, Analyzing and Predicting Socially-Driven Image Interestingness},

Author = {Elois Berson and Ngoc Q.K. Duong and Claire-H\'{e}l\`{e}ne Demarty}, Booktitle = {Proc. of the EUSIPCO conference, Spain}, Year = {2019},

LaFin Dataset

Automatic extraction of face tracks is a key component of systems that analyzes people in audio-visual content such as TV programs and movies. Due to the lack of annotated content of this type, popular algorithms for extracting face tracks have not been fully assessed in the literature. To help fill this gap, we introduce a new dataset, based on the full audio-visual person annotation of a feature movie.

Thanks to this dataset, state-of-art tracking metrics such as track purity, can now be exploited to evaluate face tracks used by, e.g., automatic character naming systems. Also, due to consistent labeling, algorithms that aim at clustering faces or face tracks in an unsupervised fashion can benefit from this test-bed. Finally, thanks to the availability of the corresponding audio annotation, the dataset can be, e.g., used for evaluation of speaker diarization methods, and more generally for assessing multimodal people clustering or naming systems.

In order to get the data, you are asked to supply your name and email address. You will receive instructions on how to download the dataset via this email address. We may store the data you supplied in order to contact you later about benchmark related matters. The data will not be used in any other way.

To download the Hannah dataset, please send an email to 

This work is supported by AXES EU project.

A. Ozerov, J.-R. Vigouroux, L. Chevallier and P. Pérez. On evaluating face tracks in movies. In Proc. Int. Conf. Image Proc. (ICIP), 2013.

Hannah Dataset

The Interestingness Dataset is a collection of movie excerpts and key-frames and their corresponding ground-truth files based on the classification into interesting and non-interesting samples. It is intended to be used for assessing the quality of methods for predicting the interestingness of multimedia content. The data has been produced by the 

MediaEval 2016 Predicting Interestingness

MediaEval 2017 Predicting Interestingness

 Tasks' organizers and was used in the context of this benchmark. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

 and their organizers for their support in the creation of this dataset. We also would like to thank the different co-organizers and advisories for the task:

 and Baohan Xu for their participation in computing several low-level features that are delivered with the Interestingness Dataset;

 and provider of some facial features delivered with the Interestingness Dataset;

If you make use of the Interestingness Dataset or refer to its results, please use the following citations:

M.G. Constantin, L.-D. Ștefan, B. Ionescu, Q.-K.-N. Duong, C.-H. Demarty, M. Sjoberg "Visual Interestingness Prediction: A Benchmark Framework and Literature Review", International Journal of Computer Vision, ISI Impact Factor 5.7, 2021.

Title = {Visual Interestingness Prediction: A Benchmark Framework and Literature Review},

Author = {Mihael Constanting and Liviu Stefan and Bogdan Ionescu and Ngoc Duong and Claire-H\'{e}l\`{e}ne Demarty and Mats Sj\¨{o}berg},

Booktitle = {nternational Journal of Computer Vision},

C.-H. Demarty, M. Sjöberg, B. Ionescu, T.-T. Do, M. Gygli and N. Q. Duong. Mediaeval 2017 predicting media interestingness task. In Proc. Of the MediaEval 2017 Workshop, Dublin, Ireland, September 13-15, 2017 (

Title = {MediaEval 2017 Predicting Media Interestingness Task},

Author = {Claire-H\'{e}l\`{e}ne Demarty and Mats Sj\¨{o}berg and Bogdan Ionescu and Than-Toan Do and Mychael Gygli and Ngoc Q.K. Duong},

Booktitle = {Proc. of the MediaEval 2017 Workshop, Dublin, Ireland, Sept. 13-15},

C.-H. Demarty, M. Sjöberg, B. Ionescu, T.-T. Do, H. Wang, N. Q. Duong, and F. Lefebvre. Mediaeval 2016 predicting media interestingness task. In Proc. Of the MediaEval 2016 Workshop, Hilversum, Netherlands, Oct. 20-21, 2016 (

Title = {MediaEval 2016 Predicting Media Interestingness Task},

Author = {Claire-H\'{e}l\`{e}ne Demarty and Mats Sj\¨{o}berg and Bogdan Ionescu and Than-Toan Do and Hanli Wang and Ngoc Q.K. Duong and Fr\'{e}d\'{e}rique Lefebvre},

Booktitle = {Proc. of the MediaEval 2016 Workshop, Hilversum, Netherlands, Oct. 20-21},

Interestingness Dataset

The recent VR/AR applications require some way to evaluate the Quality of Experience (QoE) which can be described in terms of comfort, acceptability, realism, and ease of use. In order to assess all these different dimensions, it is necessary to take into account the user’s senses and in particular, for A/V content, the vision. Understanding how users watch a 360° image, how they scan the content, where they look and when is thus necessary to develop appropriate rendering devices and create good content for consumers. The goal of this database is to help solve those questions and design good models.

The dataset contains 60 images (360°), along with eye-tracking data provided as scan-path and saliency maps and collected from 48 different observers. The data has been produced by Technicolor and by the University of Nantes (LS2N laboratory) in 2017. It has been described in several publications.


We would like to thank the different users who participated in this capture.
The creation of this benchmark has also been supported, in part, by: 

, a Marie-Curie ITN of the European Commission.


Please cite the following paper in your publications making use of the 

Yashas Rai, Patrick Le Callet and Philippe Guillotel. “Which saliency weighting for omnidirectional image quality assessment?”. In 

Proceedings of the IEEE Ninth International Conference on Quality of Multimedia Experience

 (QoMEX’17). Erfurt, Germany, pp. 1-6, June 2017.

Yashas Rai, Jesús Gutiérrez, and Patrick Le Callet. 2017. “A Dataset of Head and Eye Movements for 360 Degree Images“. In 

Proceedings of the 8th ACM on Multimedia Systems Conference

 (MMSys'17). ACM, New York, NY, USA, pp. 205-210, June 2017.

Salient 360! Dataset

is a collection of ground-truth files based on the extraction of violent events in movies and web videos, together with high-level audio and video concepts. It is intended to be used for assessing the quality of methods for the detection of violent scenes and/or the recognition of some high level, violence-related, concepts in movies and web videos.

The data was produced by Technicolor for the 2012 subset and by the 

 for the 2013 and 2014 subsets. It has been described in several publications. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

This dataset was used in the multimodal benchmark 

We would like to thank the MediaEval benchmark and their organizers for their support in the creation of this dataset. We also would like to thank the different co-organizers during all these past years:

, co-organizers of the Affect Task in 2011 and 2012;

, co-organizers of the Affect Task in 2013 and 2014, who also participated in the definition of the violent segments;

, who joined and led the organizers team in 2014;

The creation of this benchmark has also been supported, in part, by:

the China’s National 973 Program (#2010CB327900)

the China’s NSF Projects (#61201387 and #61228205)

Academy of Finland funding grants no. 255745 and 251170.

UEFISCDI SCOUTER (under grant no. 28DPST/30-08-2013).

EU FP7-ICT-2011-9: project no. 601166 ("PHENICX")

If you make use of the VSD dataset, or refer to its results, please use the following citations:

, C. Penet, M. Soleymani, G. Gravier. VSD, a public dataset for the detection of violent scenes in movies: design, annotation, analysis and evaluation. In 

, B. Ionescu, Y.G. Jiang, and C. Penet. Benchmarking Violent Scenes Detection in movies. In Proceedings of the 2014 12th International Workshop on Content-Based Multimedia Indexing (CBMI), 2014. (

M. Sjöberg, B. Ionescu, Y.G. Jiang, V.L. Quang, M. Schedl and 

. The MediaEval 2014 Affect Task: Violent Scenes Detection. In Working Notes Proceedings of the MediaEval 2014 Workshop, Barcelona, Spain (2014). (

, C. Penet, G. Gravier and M. Soleymani. A benchmarking campaign for the multimodal detection of violent scenes in movies. 

In Proceedings of the 12thinternational conference on Computer Vision – Volume Part III (ECCV’12), 

Andrea Fusiello, Vittorio Murino, and Rita Cucchiara (Eds), Col. Part III. Springer Verlag, Berlin. (

Markus Schedl, Mats Sjöberg, Ionuţ Mironică, Bogdan Ionescu, Vu Lam Quang, Yu-Gang Jiang, 

. VSD2014: A dataset for violent scenes detection in Hollywood movies and web videos. CBMI 2015

Violent Scenes Dataset

The Movie Memorability Database and related software is a collection of movie excerpts and corresponding ground-truth files based on the measurement of long-term memory performance when recognizing small movie excerpts from weeks to years after having viewed them. It is accompanied with audio and video features extracted from the movie excerpts.

It is intended to be used for assessing the quality of methods for predicting the memorability of multimedia content. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

All documents and papers that report on research that uses the Movie Memorability Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:


2018 R.Cohendet, K. Yadati, N. Q. Duong and C.-H. Demarty. Annotating, understanding, and predicting long-term video memorability. In Proceedings of the ICMR 2018 Conference, Yokohama, Japan, June 11-14, 2018.


Title = {Annotating, understanding, and predicting long-term video memorability},

Author = {Romain Cohendet and Karthik Yadati and Ngoc Q.K. Duong and Claire-H\'{e}l\`{e}ne Demarty},

Booktitle = {Proc. of the ICMR 2018 Workshop, Yokohama, Japan, June 11-14},

It would be highly appreciated if this use was shared with InterDigital, at this address: 

Movie Memorability Dataset

The automatic recognition of human emotions is of great interest in the context of multimedia applications and brain-computer interfaces. While users’ emotions can be assessed based on questionnaires, the results may be biased because the answers could be influenced by social expectations. More objective measures of emotions can be obtained by studying the users' physiological responses. The present database has been constructed in particular to evaluate the usefulness of electroencephalography (EEG) for emotion recognition in the context of audio-visual stimuli, but it also contains simultaneous physiological recordings (electrocardiogram, respiration, blood oxygen level, pulse rate, galvanic skin response) in addition to the EEG data. To the best of our knowledge, this database is the first publicly available one containing high-resolution (HR) EEG recordings for studies of emotion.

In order to study the emotions elicited by audio-visual stimuli (film excerpts), we recruited 40 subjects to participate in an experiment. During the experiment, we presented 13 emotional videos (7 videos for positive emotions and 6 videos for negative emotions) and 13 neutral videos to the subjects. The subjects were comfortably installed on a chair at about 1m distance from the 21'' computer which was used for the presentation of the stimuli.

The experiment was implemented in E-Prime 2.0 (Psychology Software Tools, Pittsburgh, PA) according to the experimental protocol shown in Figure 1. Except for the test trial, both the neutral videos and the emotional videos were presented in an arbitrary order. The emotional stimuli were selected from the FilmStim database (A. Schaefer, F. Nils, X. Sanchez, and P. Philippot, “Assessing the effectiveness of a large database of emotion-eliciting films: a new tool for emotion researchers,” Cognition & Emotion, vol. 24, no. 7, pp. 1153–1172, 2010, http://nemo.psp.ucl.ac.be/FilmStim/) and were between 40 s and 6 min long. After each emotional video, the subjects evaluated the emotions felt during the video on a negative (-1) - neutral (0) - positive (1) discrete scale. HR-EEG recordings and physiological measurements (1-channel electrocardiogram, respiration, blood oxygen level, pulse rate) were acquired with a 257-channel EGI system (EGI, Electrical Geodesics Inc., Eugene, USA) at a sampling rate of 1000 Hz and galvanic skin response measurements were recorded simultaneously using a SenseWear sensor (BodyMedia Inc.) with a sampling rate of 31 Hz. The experiment took place in a shielded room in Pontchaillou hospital, Rennes, France, and was approved by the local ethical review board of Inserm.

The database includes the EEG and other physiological recordings of the 40 subjects collected during the viewing of neutral and emotional videos and for the black screen periods. The data are provided in Matlab file format. Further information on the subjects (age and gender), the individual ratings of the videos (self-assessment labels), and the experiment (order of presentation of the videos) is also included in the Matlab files, along with all the necessary details on preprocessing and data analysis required to reproduce the results presented in (H. Becker, J. Fleureau, P. Guillotel, F. Wendling, I. Merlet, and L. Albera, “Emotion recognition based on high-resolution EEG recordings and reconstructed brain sources”, submitted to IEEE Transactions on Affective Computing, 2016). Moreover, we provide a Matlab script to read this data. More details about the data and the contents of the Matlab files are 

Please cite the following paper in your publications making use of the HR-EEG4EMO database:


H. Becker, J. Fleureau, P. Guillotel, F. Wendling, I. Merlet, and L. Albera, “Emotion recognition based on high-resolution EEG recordings and reconstructed brain sources”, submitted to IEEE Transactions on Affective Computing, 2016

HR-EEG4EMO Dataset

We provide a set of synchronized Light-Field video sequences captured by a 4x4 camera rig at 30fps. Each camera has a resolution of 2048x1088 pixels and a 12mm lens. The Field Of View (FOV) is 50˚ x 37˚.

For each Light-Field video sequence we provide the captured images after the color homogenization and the demosaicking, as well as the pseudo-rectified images. Images are provided in PNG format. Calibration parameters are also provided. Please see [1] for more details on how to use our dataset.

CITING THE INTERDIGITAL LIGHT-FIELD DATASET

 Dataset and Pipeline for Multi-View Light-Field Video

. N. Sabater, G. Boisson, B. Vandame, P. Kerbiriou, F. Babon, M. Hog, T. Langlois, R. Gendrot, O. Bureller, A. Schubert, V. Allie. CVPR Workshops, 2017.

Dataset and Pipeline for Multi-View Light-Field Video

Author = {Sabater, Neus and Boisson, Guillaume and Vandame, Benoit and Kerbiriou, Paul and Babon, Frederic and Hog, Matthieu and Langlois, Tristan and Gendrot, Remy and Bureller, Olivier and Schubert, Arno, and Allie, Valerie},

Light-Field Dataset

The VideoMem or Video Memorability Database is a collection of sound-less video excerpts and their corresponding ground-truth memorability files. The memorability scores are computed based on the measurement of short-term and long-term memory performances when recognizing small video excerpts a few minutes after viewing them for the short-term case, and 24 to 72 hours later, for the long-term case. It is accompanied with video features extracted from the video excerpts.

It is intended to be used for understanding the memorability of videos and for assessing the quality of methods for predicting the memorability of multimedia content. A detailed description of the benchmark can be found on our Data Description page. The license conditions are mentioned on the Download page.

All documents and papers that report on research that uses the VideoMem Dataset must acknowledge the use of the dataset by including an appropriate citation to the following:


R. Cohendet, C.-H. Demarty, N. Q. Duong and M. Engilberge. VideoMem: Constructing, Analyzing, Predicting Short-term and Long-term Video Memorability. ICCV 2019.


Title = {VideoMem: Constructing, Analyzing, Predicting Short-term and Long-term Video Memorability},

Author = {Romain Cohendet and Claire-H\'{e}l\`{e}ne Demarty and Ngoc Q.K. Duong and Martin Engilberge},

Booktitle = {Proceedings of the International Conference on Computer Vision, Seoul},


R. Cohendet, C.-H. Demarty, N. Q. Duong and M. Engilberge. VideoMem: Constructing, Analyzing, Predicting Short-term and Long-term Video Memorability. 2018. Arxiv:1812.01973.


, with “VideoMem” in the subject of your email.

DATA SETS