shihpolee/ESTANet_datasets

ESTANet Online Error Detection Datasets

This repository provides processed frame-level annotations and pre-extracted visual features for online error detection in procedural videos, as used in:

ESTANet: Efficient Online Error Detection in Procedural Videos via Prediction Inconsistency, ECCV 2026.

Paper: https://arxiv.org/abs/2606.25317

The release covers three benchmark datasets:

• EgoPER
• EPIC-Tent-O
• Assembly101-O

The data is intended for online error detection and temporal action segmentation experiments, where models observe procedural videos over time and predict whether the ongoing execution is correct or erroneous.

Dataset Details

Dataset Description

The dataset package contains per-frame action labels, per-frame error labels where available, and pre-extracted video features. It is designed to support the ESTANet evaluation protocol, which detects procedural errors from prediction inconsistencies among action detectors with different temporal contexts and sensitivity.

Dataset Sources

• Repository / codebase: ESTANet
• Paper: https://arxiv.org/abs/2606.25317
• arXiv DOI: https://doi.org/10.48550/arXiv.2606.25317

Dataset Variants

Dataset	Subsets / Tasks	Feature Type	Target Format
EgoPER	coffee, oatmeal, pinwheels, quesadilla, tea	vc_v_features	target_perframe/*.npy, framewise_action_error_labels/*.txt
EPIC-Tent-O	tent assembly	rgb_anet_resnet50	target_perframe/*.npy
Assembly101-O	assembly procedures	rgb_anet_resnet50	target_perframe/*.npy

Repository Structure

Expected layout:

EgoPER/
  action2idx.json
  idx2action.json
  coffee/
    target_perframe/
    framewise_action_error_labels/
    vc_v_features/
  oatmeal/
    target_perframe/
    framewise_action_error_labels/
    vc_v_features/
  pinwheels/
    target_perframe/
    framewise_action_error_labels/
    vc_v_features/
  quesadilla/
    target_perframe/
    framewise_action_error_labels/
    vc_v_features/
  tea/
    target_perframe/
    framewise_action_error_labels/
    vc_v_features/

PREGO/
  Epic-tent-O/
    target_perframe/
    rgb_anet_resnet50/
  Assembly101-O/
    target_perframe/
    rgb_anet_resnet50/

Data Fields

Feature Files

Feature files are NumPy arrays saved as .npy.

• EgoPER visual features: shape (T, 256), dtype typically float32
• EPIC-Tent-O visual features: shape (T, 2048), dtype typically float32
• Assembly101-O visual features: shape (T, 2048), dtype typically float32

Here, T is the number of frames or sampled time steps for a video.

Action Target Files

Action targets are NumPy arrays saved as:

target_perframe/<video_id>.npy

Each target array has shape (T, C), where C is the number of action classes for the corresponding dataset/task. The targets are frame-level one-hot action labels.

Error Label Files

For EgoPER, framewise error labels are stored as text files:

framewise_action_error_labels/<video_id>.txt

Each line corresponds to one frame or time step and follows:

<action_label>|<execution_status>

where <execution_status> is Normal for correct execution frames and another status value for error frames.

Splits

The official train/test splits used by ESTANet are defined in data_info/video_list.json in the ESTANet codebase.

Dataset / Task	Classes	Train Videos	Test Videos
EgoPER-coffee	16	27	38
EgoPER-oatmeal	16	26	35
EgoPER-pinwheels	14	26	45
EgoPER-quesadilla	9	26	35
EgoPER-tea	11	26	35
EPIC-Tent-O	12	13	15
Assembly101-O	86	135	182

Citation

If you use this dataset or the ESTANet benchmark setup, please cite:

@inproceedings{lee2026estanet,
  title = {ESTANet: Efficient Online Error Detection in Procedural Videos via Prediction Inconsistency},
  author = {Lee, Shih-Po and Ghoddoosian, Reza and Siddiqui, Faizan and Sachdeva, Enna and Dariush, Behzad},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year = {2026},
  url = {https://arxiv.org/abs/2606.25317},
  doi = {10.48550/arXiv.2606.25317}
}