mistakeattribution/MATT-Bench

Mistake Attribution: Fine-Grained Mistake Understanding in Egocentric Videos

CVPR 2026

Yayuan Li¹, Aadit Jain¹, Filippos Bellos¹, Jason J. Corso^1,2

¹University of Michigan, ²Voxel51

[Paper] [Code] [Project Page]

---

MATT-Bench Overview

MATT-Bench provides large-scale benchmarks for Mistake Attribution (MATT) — a task that goes beyond binary mistake detection to attribute what semantic role was violated, when the mistake became irreversible (Point-of-No-Return), and where the mistake occurred in the frame.

The benchmarks are constructed by MisEngine, a data engine that automatically creates mistake samples with attribution-rich annotations from existing egocentric action datasets:

Dataset	Samples	Instruction Texts	Semantic	Temporal	Spatial
Ego4D-M	220,800	19,467	✓	✓	✓
EPIC-KITCHENS-M	299,715	12,283	✓	—	—

These are at least two orders of magnitude larger than any existing mistake dataset. Instruction-text counts = unique (predicate V, argument ARG1) pairs.

A third source, HoloAssist-M, is released alongside as an additional benchmark — see Extended: HoloAssist-M below.

Repository Layout

MATT-Bench/
├── ego4d/
│   ├── train.xlsx, valid.xlsx, test.xlsx   ← primary annotation files (consumed by the MATT codebase)
│   ├── parquet.xlsx                        ← MisEngine reproduction data (Ego4D narrations with SRL)
│   └── parquet/                            ← Parquet mirror for the HF dataset viewer
├── epickitchens/
│   ├── train.xlsx, validation.xlsx
│   └── parquet/
└── holoassist/
    ├── train.xlsx, validation.xlsx
    └── parquet/

.xlsx is the canonical download format (the MATT codebase reads Excel directly). The parquet/ mirror powers the HF dataset viewer and datasets.load_dataset(...) loaders — both views contain the same rows.

Downloading MATT-Bench

MATT-Bench has two parts that you obtain separately:

1. Annotations — semantic attribution annotations are hosted here, download via hf or git clone. Temporal and spatial attribution annotations are inherited from the original dataset.
2. Video media — not hosted here. Download from each source dataset using the instructions below. Original videos retain their upstream licenses.

Annotations (this repo)

# Everything
hf download mistakeattribution/MATT-Bench --repo-type dataset --local-dir MATT-Bench

# Just one source dataset's xlsx files
hf download mistakeattribution/MATT-Bench --repo-type dataset \
  --include "ego4d/*.xlsx" --local-dir MATT-Bench

Or via the datasets library (reads the parquet mirror):

from datasets import load_dataset
ego4d_m = load_dataset("mistakeattribution/MATT-Bench", "ego4d")
epic_m  = load_dataset("mistakeattribution/MATT-Bench", "epickitchens")
holo_m  = load_dataset("mistakeattribution/MATT-Bench", "holoassist")

Video media

Ego4D

Follow https://ego4d-data.org/docs/CLI/ to download. The video_uid and clip1_uid fields in our annotations correspond to Ego4D's native video and clip UIDs.

MATT-Bench uses the FHO (Forecasting Hands and Objects) benchmark clips from Ego4D. Example downloading script:

ego4d --output_directory="~/ego4d_data" --datasets clips --benchmarks FHO

EPIC-KITCHENS-100

Follow https://epic-kitchens.github.io/ to download. MATT-Bench's video_id matches EPIC's participant-video identifier (e.g. P22_16); start_frame / end_frame index the RGB frame sequence.

Example download script:

git clone https://github.com/epic-kitchens/epic-kitchens-download-scripts
cd epic-kitchens-download-scripts
python epic_downloader.py --rgb-frames    # or --videos

HoloAssist

Although not reported in the paper, we also support the HoloAssist dataset.

Download the following from the HoloAssist project page:

Resource	Link	Size
Videos (pitch-shifted)	video_pitch_shifted.tar	184.20 GB
Labels	data-annotation-trainval-v1_1.json	111 MB
Dataset splits	data-splits-v1_2.zip	—

MATT-Bench's video_id matches HoloAssist's video identifier (e.g. R076-21July-DSLR).

Data Schema

ego4d/{train,valid,test}.xlsx — 13 columns

Column	Description
video_uid	Ego4D video UID (full video)
start_frame, end_frame	Frame bounds of the attempt clip
clip1_uid, clip1_start_frame, clip1_end_frame	Primary Ego4D clip
clip2_uid, clip2_start_frame, clip2_end_frame	Some actions are distributed across two clips (Not required / -1 when absent)
V, ARG1	Predicate and argument from the instruction (e.g. pick up, apple)
label	Mistake label. 0: Correct; 1: Mistaken Predicate; 2: Mistaken Object; 3: Mistaken Both
split	dataset split identifier

ego4d/parquet.xlsx — 29 columns (MisEngine reproduction data)

Ego4D narration-level records with semantic-role labels (ARG0, V, ARG1), frame/time bounds (start_frame/end_frame/start_sec/end_sec), clip-relative bounds, and noun/verb embedding vectors. Used to reproduce the MisEngine step that produces the split files above.

epickitchens/{train,validation}.xlsx and holoassist/{train,validation}.xlsx — 8 columns

Column	Description
video_id	Source-dataset video identifier
start_frame, end_frame	Frame bounds of the attempt clip
V, ARG1	Predicate and argument of the instruction text
label	Mistake label
actual_V, actual_ARG1	Predicate/argument of the action performed in the video

Extended: HoloAssist-M

HoloAssist-M is an additional MATT benchmark released alongside MATT-Bench. It is not part of the main two-dataset evaluation reported in the CVPR 2026 paper; it uses the same MisEngine pipeline applied to the HoloAssist dataset.

Dataset	Samples	Instruction Texts	Semantic	Temporal	Spatial
HoloAssist-M	562,209	1,786	✓	—	—

Schema matches EPIC-KITCHENS-M (semantic attribution only — HoloAssist does not provide native PNR frame number andb bbox annotations).

Citation

@inproceedings{li2026mistakeattribution,
  title     = {Mistake Attribution: Fine-Grained Mistake Understanding in Egocentric Videos},
  author    = {Li, Yayuan and Jain, Aadit and Bellos, Filippos and Corso, Jason J.},
  booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
}

Please also cite the source datasets:

@inproceedings{grauman2022ego4d,
  title     = {Ego4D: Around the World in 3,000 Hours of Egocentric Video},
  author    = {Grauman, Kristen and others},
  booktitle = {CVPR},
  year      = {2022}
}

@article{Damen2022RESCALING,
  title     = {Rescaling Egocentric Vision: Collection, Pipeline and Challenges for EPIC-KITCHENS-100},
  author    = {Damen, Dima and others},
  journal   = {IJCV},
  year      = {2022}
}

@inproceedings{wang2023holoassist,
  title     = {HoloAssist: an Egocentric Human Interaction Dataset for Interactive AI Assistants in the Real World},
  author    = {Wang, Xin and others},
  booktitle = {ICCV},
  year      = {2023}
}