embryo-video-eval/bovine-embryo-video-evaluation

BovEmbryo: Bovine Embryo Time-Lapse Video Evaluation Dataset

Dataset Summary

BovEmbryo is a bovine embryo time-lapse video dataset for evaluating full-development biological video understanding models. The dataset contains 344 annotated videos and supports three evaluation settings: six-class IVF/SCNT outcome classification, three-class outcome-only classification, and IVF-to-SCNT / SCNT-to-IVF embryo-origin generalization.

Each video is an exported time-lapse imaging recording summarizing approximately 200 hours of embryo development. The digital playback duration is typically around 1--2 minutes and should not be interpreted as biological development duration.

BovEmbryo is released for non-commercial academic research under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.

Dataset Statistics

Six-Class Distribution (Task 1)

Class	Embryo Type	Outcome	Count
IVF-degenerate	IVF	Degenerate	53
IVF-arrested	IVF	Arrested	26
IVF-success	IVF	Success	106
SCNT-degenerate	SCNT	Degenerate	56
SCNT-arrested	SCNT	Arrested	27
SCNT-success	SCNT	Success	76
Total			344

Three-Class Distribution (Task 2 and Task 3)

Outcome	Count
Degenerate	109
Arrested	53
Success	182
Total	344

Embryo Type Distribution (Task 3)

Embryo Type	Count
IVF	185
SCNT	159
Total	344

Evaluation Tasks

Task 1: Six-Class IVF/SCNT Outcome Classification

Classify each video into one of six categories combining embryo type and developmental outcome:

• IVF-degenerate
• IVF-arrested
• IVF-success
• SCNT-degenerate
• SCNT-arrested
• SCNT-success

Task 2: Three-Class Outcome-Only Classification

Classify the embryo developmental outcome independent of embryo production type:

• degenerate
• arrested
• success

Task 3: IVF ↔ SCNT Embryo-Origin Generalization

Evaluate cross-domain generalization using embryo production type as the domain variable:

• Train and validate on IVF; test on SCNT
• Train and validate on SCNT; test on IVF

Baseline Results

Task 1 and Task 2 results are reported as mean ± standard deviation over stratified 5-fold cross-validation. Task 3 results are reported per transfer direction using a fixed source-origin / target-origin protocol. Full results across all frame settings and model configurations are reported in the accompanying paper.

Task 1: Six-Class IVF/SCNT Outcome Classification

Model	Frames	Acc (%)	Macro F1 (%)	Macro Precision (%)	Macro Recall (%)
ResNet152 + BiLSTM-256	32	66.86 ± 3.57	60.87 ± 5.49	65.37 ± 7.48	62.10 ± 5.08
ResNet152 + BiLSTM-256	128	66.56 ± 3.64	62.55 ± 9.03	61.09 ± 3.70	59.24 ± 5.76
SlowFast R50	Fast 32 / Slow 8	67.43 ± 3.12	51.52 ± 3.44	51.64 ± 5.35	54.12 ± 3.05
SlowFast R50	Fast 128 / Slow 32	65.98 ± 5.64	51.35 ± 10.60	53.24 ± 15.72	53.55 ± 7.87
Video Swin Transformer	32	75.02 ± 6.12	67.75 ± 5.13	76.47 ± 7.94	67.76 ± 4.52
Video Swin Transformer	128	75.87 ± 3.72	68.39 ± 5.06	71.56 ± 5.31	68.30 ± 4.54
Video Swin Transformer	256	76.16 ± 5.58	69.67 ± 7.09	75.18 ± 7.54	69.58 ± 7.18
VideoMAE Base	16	79.06 ± 5.29	71.66 ± 7.91	72.52 ± 8.34	71.86 ± 7.60
VideoMamba-Middle	64	56.75 ± 11.23	44.47 ± 16.99	45.64 ± 19.75	47.02 ± 13.90

Task 2: Three-Class Outcome-Only Classification

Model	Frames	Acc (%)	Macro F1 (%)	Macro Precision (%)	Macro Recall (%)
ResNet152 + BiLSTM-256	128	77.34 ± 3.41	69.54 ± 4.48	73.27 ± 3.89	71.47 ± 5.02
SlowFast R50	Fast 32 / Slow 8	84.87 ± 2.07	75.83 ± 3.62	80.68 ± 3.79	74.98 ± 2.90
Video Swin Transformer	256	88.96 ± 3.25	83.97 ± 4.88	85.98 ± 5.26	82.91 ± 4.89
VideoMAE Base	16	86.64 ± 3.20	78.79 ± 7.25	82.19 ± 7.28	78.12 ± 7.13
VideoMamba-Middle	64	57.55 ± 3.47	36.92 ± 10.30	38.82 ± 13.72	41.33 ± 6.51

Task 3: IVF ↔ SCNT Embryo-Origin Generalization

Results are reported per transfer direction. Target columns show performance on the unseen target origin.

Model	Direction	Target Acc (%)	Target F1 (%)	Target Precision (%)	Target Recall (%)
ResNet152 + BiLSTM-256	IVF → SCNT	74.84	66.39	68.41	65.79
ResNet152 + BiLSTM-256	SCNT → IVF	67.57	60.33	61.38	62.53
SlowFast R50	IVF → SCNT	55.97	46.31	48.84	47.27
SlowFast R50	SCNT → IVF	74.05	59.52	62.39	59.87
Video Swin Transformer	IVF → SCNT	71.07	56.13	57.70	57.59
Video Swin Transformer	SCNT → IVF	63.24	52.33	54.47	54.06
VideoMAE Base	IVF → SCNT	74.84	67.16	68.06	67.55
VideoMAE Base	SCNT → IVF	75.68	62.55	65.91	61.18
VideoMamba-Middle	IVF → SCNT	46.54	31.24	29.62	34.81
VideoMamba-Middle	SCNT → IVF	50.81	31.21	33.98	33.38

Dataset Structure

The repository is organized as follows:

data/
  videos/
    1-IVF-degenerate/
    2-IVF-arrested/
    3-IVF-success/
    4-SCNT-degenerate/
    5-SCNT-arrested/
    6-SCNT-success/
metadata/
  videos_metadata.csv
  labels_6class.csv
  labels_3class.csv
  class_distribution.csv
  outcome_distribution.csv
  embryo_type_distribution.csv
  embryo_type_outcome_distribution.csv
  dataset_summary.json
  class_mapping.json
docs/
  sample/
    1-IVF-degenerate/
    2-IVF-arrested/
    3-IVF-success/
    4-SCNT-degenerate/
    5-SCNT-arrested/
    6-SCNT-success/
  DATA_CARD.md
  EVALUATION_CARD.md
  annotation_protocol.md
croissant.json

Sample Videos

The files under docs/sample/ are provided only for quick visual inspection by reviewers and users. They are not part of the 344-video benchmark dataset and should not be included in training, validation, testing, or reported dataset statistics. The official benchmark records are the 344 videos listed in metadata/videos_metadata.csv.

Metadata Files

• videos_metadata.csv: master metadata file with one row per video, including video ID, file path, embryo type, outcome labels, video properties, playback duration, approximate biological duration, and file size.
• labels_6class.csv: label file for Task 1, six-class IVF/SCNT outcome classification.
• labels_3class.csv: label file for Task 2 and Task 3, three-class outcome-only classification and embryo-origin generalization.
• class_distribution.csv: six-class count and percentage distribution.
• outcome_distribution.csv: three-class outcome distribution.
• embryo_type_distribution.csv: IVF and SCNT embryo type distribution.
• embryo_type_outcome_distribution.csv: embryo-type-by-outcome distribution for Task 3 context.
• dataset_summary.json: dataset-level summary statistics.
• class_mapping.json: mapping between integer labels, class names, embryo types, and domain-shift protocols.
• croissant.json: Croissant machine-readable metadata file with core metadata and Responsible AI metadata.

Intended Use

This dataset is intended for non-commercial academic research on biological video understanding, embryo viability assessment, temporal representation learning, and benchmarking of video classification models under full-trajectory, low-shot, and domain-shift conditions.

Out-of-Scope Use

This dataset is not intended for direct clinical, veterinary, reproductive, or commercial decision-making without additional validation, regulatory review, and explicit authorization. It should not be used for commercial embryo ranking, direct embryo transfer decisions, pregnancy or live-birth inference, or unsupported deployment outside the validated benchmark setting.

Limitations

The dataset is limited in scale and originates from a controlled laboratory imaging setting. Performance on this dataset should not be interpreted as evidence of generalization across laboratories, imaging systems, species, culture protocols, or deployment environments. Class imbalance is present, particularly in the arrested class. Some classes are visually similar, especially degenerate and arrested embryos, and visual differences between IVF and SCNT embryos may be subtle.

License

BovEmbryo is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license for non-commercial academic research use. Commercial use is not permitted under this license.