Access to the Infant Care Video Dataset requires approval and agreement to the ICVD Data Use Agreement version 2.0, dated 10 September 2026, in LICENSE.md. Use is limited to non-commercial research, teaching and study as defined there. You may share access with members of your research group working under your supervision who agree to those terms; anyone else must request access individually. Other sharing is limited to the agreement's permissions for derived materials, illustrative samples and public documentation. Resulting publications and releases must acknowledge the dataset version and cite doi 10.1109/COMPSAC65507.2025.00299. An additional copy of the paper is available at arXiv:2608.23838.

Infant Care Video Dataset (ICVD) v2.0

4,144 short videos of twelve simulated infant care interventions, recorded on a manikin with eight cameras, for training and evaluating video classifiers aimed at automating clinical documentation.

Project page. infantcaredataset.org

Paper. Igor Bogdanov and James Green, "Infant Care Video Dataset for Classification of Interventions Using Transformers", IEEE COMPSAC 2025, pages 2130 to 2135.

Published version: doi 10.1109/COMPSAC65507.2025.00299
An additional copy of the paper is available at arXiv:2608.23838.

Abstract

The following abstract is reproduced from the paper. Its performance figures refer to the files used in that study; results measured on v2 are reported in Benchmarks.

Healthcare documentation in the neonatal intensive care unit (NICU) presents significant challenges, with nurses spending approximately 25% of their time on record-keeping, while up to 60% of interventions remain undocumented. Motivated by the need to detect interventions from video automatically, we present the Infant Care Video Dataset (ICVD), a collection of 4,144 videos spanning 12 simulated intervention classes designed for developing automated documentation systems. Our manikin-based approach systematically varies conditions, such as camera angle and clinician skin tone, while ensuring privacy compliance. Using video transformer architectures (TimeSformer and MotionFormer), we establish strong baseline performance (93.97% and 93.17% top-1 accuracy) among the 12 infant care classes. Our ablation study comparing temporal models with a framewise approach (23.17% accuracy) demonstrates a 70.80% performance gap, validating the need for temporal modeling. The ICVD provides a foundation for developing automated documentation systems to reduce clinical burden in neonatal care environments and improve existing practices.

Two resolutions of the same 4,144 clips, with the same identifiers and splits. The clips contain 18.54 hours of video in total; the two resolutions are alternative encodings of that same material.

Folder	Resolution	Size	Use
`456x256/`	456 x 256	3.4 GB	v2 benchmarks using the paper's training protocols
`640x360/`	640 x 360	9.0 GB	applications requiring greater spatial detail

Each folder holds one archive per split plus the tables loose, so the tables can be read without downloading the video.

456x256/
  icvd-456x256-train.tar     2.40 GB
  icvd-456x256-val.tar       0.52 GB
  icvd-456x256-test.tar      0.53 GB
  train.csv val.csv test.csv classes.csv metadata.csv CHECKSUMS.sha256

Inside an archive: <split>/<class>/<background>/<id>.mov.

Extracting the videos

The release includes extract.py, a small helper using only the Python standard library (Python 3.8 or newer). After access is approved, download the root files and all files in the desired resolution folder, keeping the repository layout. Each selected resolution needs its three archives and accompanying tables.

Run from the directory containing extract.py:

python3 extract.py

This verifies the archive checksums and extracts all downloaded resolutions into a new extracted/ directory. Each resolution contains the train/, val/ and test/ video folders together with its CSV tables and clip checksums. The Data Use Agreement and citation are copied to the output root. The downloaded archives are retained.

To extract one resolution into a different directory:

python3 extract.py --resolution 456x256 --output ./icvd-456

The output directory must not already exist. The helper refuses to overwrite existing data, and an unsuccessful extraction removes its temporary output. Unpacked videos require additional disk space.

Splits and classes

Split	Clips
train	2,891
val	623
test	630

Index	Class	Clips
0	wiping-body	610
1	crib-take	337
2	blanket-cover	319
3	diaper-put-on	310
4	wiping-face	324
5	pacifier-take	321
6	feeding	315
7	pacifier-give	334
8	crib-put	339
9	blanket-uncover	315
10	diaper-change	305
11	diaper-remove	315

wiping-body has roughly double the clips because it was performed both with and without a diaper. Splits are 70 / 15 / 15, drawn separately within each class and each background so the proportions hold in every split. The split is the one used in the paper and is unchanged in v2.

Files

train.csv, val.csv and test.csv have a header row and two columns, path and class_number, followed by one row per clip. Paths are relative to the extracted resolution directory; for example, extracted/456x256/. classes.csv maps each zero-based class index to its class name.

metadata.csv adds, per clip: video id, split, class, background, hand covering, camera id, camera model, camera type, width, height, frame rate, duration, size and SHA-256. Frame rate is the nominal rate, 24, 30 or 60. Several cameras record at the broadcast forms of those rates, 24000/1001 for example, which the files themselves declare. Every clip is constant rate: the declared rate matches the average rate on all 4,144.

Clip identifiers are opaque and do not encode recording or production details. The mapping to production files is not included in the release.

Recording conditions

One session, eight cameras running in parallel, as listed in the paper: two Arenti 360 View 5G WiFi Pet Monitors, an Arenti 4MP Baby Monitor, two Apple iPhone 11 Pro Max, an Apple iPhone 8 Plus, a Google Pixel 7 Pro and a Samsung Galaxy S22 Ultra. camera_id in the metadata runs 1 to 8 and is stable across the dataset, so clips can be grouped by camera without naming any file. Systematically varied: light and dark crib sheets, five hand conditions (bare hands, and white, green, dark blue and light blue gloves, 812 to 852 clips each), natural and artificial lighting, and eight viewpoints. Nominal frame rates are 24 fps for the three Arenti monitors, 30 or 60 fps for both iPhone 11 Pro Max cameras, and 30 fps for the remaining cameras. Across the dataset, 1,572 clips are nominally 24 fps, 1,954 are 30 fps and 618 are 60 fps. Clip duration ranges from 0.87 to 66.48 seconds, with a mean of 16.10 seconds.

Changes from the version described in the paper

Video files contain no audio.
All identifying metadata is stripped from the video files.
Clip identifiers differ from those used in the paper's experiments.
Three errors in the paper's test.csv are corrected: a label typed 6c, a quoted path, and a clip with no row.
Split sizes are 2,891 / 623 / 630. The paper's Table I reports one more per split because it counted the header row of each file.
The 456x256 tier is encoded at higher quality than the paper's files. Its three archives total 3.4 GB.

The benchmarks below distinguish the paper's results on the original files from new measurements on the reprocessed v2 release.

Benchmarks

The v2 benchmark uses the 456x256 release, with 2,891 training clips, 623 validation clips and 630 test clips. Models are fine-tuned from Kinetics-400 pretrained checkpoints using the original training and evaluation scripts and the configurations saved with the successful paper checkpoints. The v2 runs execute on an Apple M5 Max through PyTorch's MPS backend, with device compatibility applied at runtime. The training and evaluation code, pretrained weights and trained checkpoints are not included in this dataset release; the table below documents the protocol used to obtain these results.

All values below are percentages. Precision, recall and F1 are macro averages across the 12 classes.

Model	Evaluation	Top-1	Top-5	Precision	Recall	F1
TimeSformer	Files used in the paper, Table III	93.97	100.00	94.11	93.29	92.92
Motionformer	Files used in the paper, Table III	93.17	99.84	93.73	92.32	91.29
TimeSformer	v2, completed	92.70	100.00	93.00	92.07	91.77
Motionformer	v2, completed	93.65	99.84	93.45	92.86	92.25

TimeSformer correctly classifies 584 of 630 test clips; Motionformer correctly classifies 590 of 630. Both runs completed their full training schedules and final multi-view evaluation. Their final-epoch checkpoint identities and per-clip predictions have been checked against the released test labels and the original evaluators' metrics.

Setting	TimeSformer	Motionformer
Frames per input / sampling rate	20 / 32	16 / 4
Input crop	224 x 224	224 x 224
Epochs / evaluated checkpoint	15 / final epoch	35 / final epoch
Optimizer	SGD, momentum 0.9, Nesterov	AdamW
Initial learning rate / weight decay	0.005 / 0.0001	0.0001 / 0.05
Learning-rate schedule	10-fold reductions at indices 11 and 14	10-fold reductions at indices 20 and 30
Training batch	8 clips; accumulate 8 batches, effective batch 64	8 clips
Test views per clip	1 temporal view x 3 spatial crops	10 temporal views x 3 spatial crops
Training precision / PyTorch	FP32 / 2.1.2	FP16 mixed precision / 2.14.0
Random seed	0	0

Sampling rates use the original loaders' 30-fps reference. Learning-rate indices are zero-based. The original sampling, augmentation, optimization and evaluation procedures are retained.

TimeSformer's v2 top-1 accuracy is 1.27 percentage points below the paper's result, and its macro F1 is 1.15 points below. Motionformer's top-1 accuracy is 0.48 percentage points above the paper's result, and its macro F1 is 0.96 points above. These are new fine-tuning runs on reprocessed videos in a different execution environment. A single seed does not establish run-to-run variability, and this comparison does not isolate the effect of individual release changes.

Limitations

The dataset contains simulated, trimmed interventions recorded on a manikin in one session. All eight cameras occur in each split. The supplied split measures clip classification within these recording conditions and does not establish generalization to new sessions, unseen cameras or real neonatal care. Glove colours provide controlled variation in hand appearance; they are not measurements of clinician skin tone. Performance on these clips also does not establish accuracy for detecting interventions in continuous clinical video.

Licence and citation

Non-commercial research, teaching and study under the Data Use Agreement, version 2.0. The agreement permits preprocessing and model training, fine-tuning and evaluation. Members of your research group working under your supervision may use your copy after agreeing to the terms; anyone else must request access individually.

Clips, extracted frames, modified copies of source imagery and the supplied data tables may not be redistributed except as permitted by the agreement. Derived models, features and statistics may be published and redistributed for non-commercial research with the required attribution and terms preserving the agreement's restrictions. Limited illustrative samples may accompany research publications, including through commercial publishers. README.md, LICENSE.md and CITATION.cff may be shared with their notices intact.

The included extract.py helper is MIT licensed, with the full notice inside the script. The dataset remains governed by the Data Use Agreement. Acknowledge the dataset version used and cite the published paper below. An additional copy of the paper is available at arXiv:2608.23838.

@inproceedings{bogdanov2025icvd,
  title       = {Infant Care Video Dataset for Classification of Interventions Using Transformers},
  author      = {Bogdanov, Igor and Green, James},
  booktitle   = {2025 IEEE 49th Annual Computers, Software, and Applications Conference (COMPSAC)},
  pages       = {2130--2135},
  year        = {2025},
  publisher   = {IEEE},
  doi         = {10.1109/COMPSAC65507.2025.00299},
  eprint      = {2608.23838},
  eprinttype  = {arXiv},
  note        = {Additional copy available at arXiv:2608.23838}
}

Contact

Igor Bogdanov, igor@isbogdanov.com, infantcaredataset.org. Access requests are reviewed manually; commercial use and any question about the terms go to the same address.

Integrity

SHA256SUMS.txt covers the archives. The extraction helper verifies these before unpacking. Each tier's CHECKSUMS.sha256 covers every clip inside. Archive hashes can also be checked manually:

sha256sum -c SHA256SUMS.txt

Downloads last month: 16

Total file size:

12.5 GB

Paper for isbogdanov/ICVD

Infant Care Video Dataset for Classification of Interventions Using Transformers

Paper • 2608.23838 • Published 26 days ago

You need to agree to share your contact information to access this dataset