JNTD-DS: A Benchmark Dataset for Just Noticeable frame rate-based Temporal Difference in Perceptual Video Coding

This page hosts the dataset and code for the paper "JNTD-DS: A Benchmark Dataset for Just Noticeable frame rate-based Temporal Difference in Perceptual Video Coding."

JNTD Database

The JNTD dataset is a high-frame-rate video dataset designed to study the Just-Noticeable Temporal Difference (JNTD) in video sequences. It contains 50 reference videos at a spatial resolution of 1920 × 1080, each 5 seconds long at 60 fps, along with 10 temporally degraded versions for each reference. These versions are generated at 10 distinct frame rates: 10, 15, 20, 25, 30, 35, 40, 45, 50, and 55 fps.

The dataset covers a wide range of scene content, including people, animals, buildings, transportation, nature, sports, objects, and computer-generated graphics. In total, the dataset includes 550 video sequences, enabling analysis of when changes in temporal resolution become perceptible to the human visual system.

Subjective test

Each participant compared the quality of videos in terms of frame rate. At each step, they compared a test video with reduced quality (lower frame rate) to an original video of the same scene at the highest quality, and reported whether they noticed the change in quality or not. Based on their response, another pair of original and test videos was displayed until the JNTD point was found.

Video sequences in the JNTD-DS

Requirements

• FFmpeg
• Tensorflow

Download

📄 jntd_dataset.csv (Subjective test results)

This repository includes a jntd_dataset.csv file containing all subjective test results used to determine the Just-Noticeable Temporal Difference (JNTD) for each video.

The CSV file includes the following fields:

Column Name	Description
Video ID	A unique index (1–50) assigned to each reference video in the JNTD dataset.
Source dataset and video name	Indicates the origin of the video (BVI-HFR, UVG, or Inter4K) along with the original sequence name.
Number of participants	Total number of viewers who rated the video in the subjective test.
Individual ratings	A list of raw user responses (e.g., [25, 40, 15, ...]), each indicating the minimum frame rate at which temporal degradation became noticeable.
Final JNTD value	The aggregated JNTD threshold (e.g., using 75% Satisfied User Ratio), representing the frame rate where temporal artifacts become perceptible.

🎥 Source Videos

The JNTD dataset was constructed by selecting videos from the following three publicly available high-frame-rate (HFR) video datasets:

• BVI-HFR: All 22 video sequences from this dataset were included.

• UVG: 7 video sequences were selected: Beauty, Bosphorus, HoneyBee, Jockey, ReadySetGo, YachtRide, and Lips.

• Inter4K: 21 video sequences were selected: 4, 6, 9, 19, 38, 43, 81, 107, 145, 267, 333, 411, 550, 561, 619, 695, 731, 821, 950, 983, and 993.

All selected source videos were processed and resampled using FFmpeg to generate multiple temporally degraded versions, producing the final JNTD dataset suitable for analyzing just-noticeable temporal differences in videos.

To support reproducibility, pre-made batch files are provided for each source dataset:

• process_BVI_HFR.bat
• process_UVG.bat
• process_Inter4K.bat

JNTD Prediction Model

The pre-trained model can be downloaded from this link and used directly for JNTD inference.
In addition, the provided code supports training a new model on a custom dataset.

Note: The dataset used for training and testing should have such a structure.

- Dataset/
        - train/
              - frame/
                     - img#1
                     - ...
              - diff/
                     - img#1
                     - ...
              - optf/
                     - img#1
                     - ...
              - labels.txt

        - valid/
              - frame/
                     - img#1
                     - ...
              - diff/
                     - img#1
                     - ...
              - optf/
                     - img#1
                     - ...
              - labels.txt

        - test/
              - frame/
                     - img#1
                     - ...
              - diff/
                     - img#1
                     - ...
              - optf/
                     - img#1
                     - ...
              - labels.txt

Dataset preparation for training and inference of the proposed model

Follow these steps to prepare the dataset:

• Use the source-video preprocessed at 60 fps from the JNTD dataset.
• Place them in the "Videos" folder, creating three subfolders: "train," "valid," and "test."
• Run "JNTD-Model/FrameExtraction.cmd" to extract the video frames using FFMPEG.
• Edit the "img_dir" and "save_dir" in "JNTD-Model/DDFlow/config/config.ini" and "JNTD-Model/DDFlow/images/sample_list.txt", then run "JNTD-Model/DDFlow/OpticalFlow.ipynb" to obtain the optical flows. For more details, you can follow the instructions.

Testing

For JNTD prediction, the following commands can be used.

python JNTD.py test --data_dir "Path-to-the-rootdir(Dataset-folder)" --pretrained_model_path "Path-to-the-pretrained-model/JNTD.h5"

Training

For JNTD training, the following commands can be used.

python JNTD.py train --data_dir "Path-to-the-rootdir(Dataset-folder)" --epochs number-of-epochs --batch_size Batch-size-for-training --learning_rate Learning-rate-for-optimizer --csv_log_path Path-to-save-CSV-logs-during-training/log.csv" --model_save_path "Path-to-the-save-model/JNTD_model.h5"

Citation

If our work is useful for your research, please kindly cite the following paper:

@article{nami2026JNTD-DS,
	title={JNTD-DS: A Benchmark Dataset for Just Noticeable frame rate-based Temporal Difference in Perceptual Video Coding},
author={Nami, Sanaz and Pakdaman, Farhad and Taali, Sahab and Hashemi, Mahmoud Reza and Shirmohammadi, Shervin and Gabbouj, Moncef},
booktitle={Proceedings of the ACM Multimedia Systems Conference},
year={2026}
}

Contact

If you have any question, leave a message here or contact Sanaz Nami (sanaz.nami@tuni.fi, snami@ut.ac.ir).