EgoExo-Next

EgoExo-Next is a visual-option benchmark for egocentric–exocentric (ego–exo) action understanding. Its prediction target is the subsequent visual state itself: rather than mapping visual evidence to an action label, a textual description, or a temporal relation, a model must identify the target state among four temporally related image candidates.

The next-state tasks deliberately include both earlier-state and later-future distractors, so the correct answer cannot be found by simply spotting the only future frame or picking the most temporally advanced candidate. Solving them requires distinguishing fine-grained progression within a single action trajectory, where candidates often share the same high-level semantics but differ in object configuration, body pose, contact relations, and spatial layout.

EgoExo-Next decomposes this capability into temporal visual-state progression and cross-view state correspondence, and then evaluates their composition. Subtask 4 in particular requires inferring an unobserved subsequent state from preceding egocentric observations and identifying how that state appears from an exocentric viewpoint — a composition that goes beyond either next-action prediction or matching already-observed ego–exo content.

• 2,503 human-curated four-choice questions
• 21,232 images, ≈6.2 GB
• 4 interconnected subtasks × 2 scene domains (indoor / outdoor)
• Built from six publicly available egocentric and ego–exo video sources

Subtasks

#	Subtask	What the model must do	Observation	Candidates
1	Ego Next-State	Temporal visual-state progression from the egocentric viewpoint	4 egocentric frames	4 next-state images
2	Ego–Exo Matching	Correspondence between synchronized ego and exo states at the same time	1 keyframe	4 images (× 2 directions)
3	Exo Next-State	Temporal visual-state progression from the exocentric viewpoint	4 exocentric frames	4 next-state images
4	Ego-to-Exo Next-State	Composition: infer the future exocentric state implied by preceding egocentric observations	4 egocentric frames	4 exocentric images

Subtask 2 is stored in both directions within the same group — ego_to_exo (egocentric keyframe → exocentric options) and exo_to_ego (exocentric keyframe → egocentric options) — each with its own observation frame, candidate set, and question text.

Contents

Directory	Subtask	Scene	Groups
Subtask1_Indoor	1 — Ego Next-State	Indoor	305
Subtask1_Outdoor	1 — Ego Next-State	Outdoor	302
Subtask2_Indoor	2 — Ego–Exo Matching	Indoor	302
Subtask2_Outdoor	2 — Ego–Exo Matching	Outdoor	302
Subtask3_Indoor	3 — Exo Next-State	Indoor	349
Subtask3_Outdoor	3 — Exo Next-State	Outdoor	316
Subtask4_Indoor	4 — Ego-to-Exo Next-State	Indoor	309
Subtask4_Outdoor	4 — Ego-to-Exo Next-State	Outdoor	318
Total			2,503

Per-subtask totals: Subtask 1 — 607, Subtask 2 — 604, Subtask 3 — 665, Subtask 4 — 627. Scene split: indoor — 1,265, outdoor — 1,238.

Repository structure

Each subtask × scene directory holds one answers.json plus one folder per question group:

Subtask1_Indoor/
├── answers.json                 # ground truth for every group in this directory
├── group0000/
│   ├── input_meta.json          # paths, question text, candidate labels
│   ├── question/                # observation frames  (1.jpg … 4.jpg)
│   └── options/                 # candidates           (A.jpg … D.jpg)
├── group0001/
└── …

Subtask 2 uses a two-direction layout — each group carries both sub-questions:

Subtask2_Indoor/
├── answers.json
├── group0000/
│   ├── input_meta.json
│   ├── ego_to_exo/
│   │   ├── question/1.jpg       # egocentric keyframe
│   │   └── options/A.jpg … D.jpg  # exocentric candidates
│   └── exo_to_ego/
│       ├── question/1.jpg       # exocentric keyframe
│       └── options/A.jpg … D.jpg  # egocentric candidates
└── …

Data format

input_meta.json — Subtasks 1, 3, 4

{
  "question": "Question:\nGiven 4 temporally ordered observation frames from an egocentric action video, which image shows the subsequent visual state following the observation sequence?\n\n…",
  "question_dir": "question",
  "option_dir": "options",
  "question_images": ["1.jpg", "2.jpg", "3.jpg", "4.jpg"],
  "options": [
    { "label": "A", "image": "A.jpg" },
    { "label": "B", "image": "B.jpg" },
    { "label": "C", "image": "C.jpg" },
    { "label": "D", "image": "D.jpg" }
  ]
}

• question is the stored question text for the group, ready to send to a model as-is. It is identical for every group within a subtask × scene directory.
• question_images are the observation frames, in temporal order, resolved against question_dir. options are the candidates, resolved against option_dir.
• Candidate entries carry only label and image — no timestamps, no ordering cues, and no indication of which candidate is correct.

input_meta.json — Subtask 2

The structural keys live inside each direction of sub_questions:

{
  "group": "group0000",
  "evaluation": "Both sub_questions must be correct for the group to pass.",
  "sample_times": [7.93, 29.1, 50.3, 71.47],
  "query_index": 1,
  "sub_questions": {
    "ego_to_exo": {
      "direction": "ego_to_exo",
      "question": "Question:\nWhich exocentric image corresponds to the egocentric image at the same time?\n\n…",
      "question_dir": "ego_to_exo/question",
      "option_dir": "ego_to_exo/options",
      "question_images": ["1.jpg"],
      "options": [
        { "label": "A", "image": "A.jpg" },
        { "label": "B", "image": "B.jpg" },
        { "label": "C", "image": "C.jpg" },
        { "label": "D", "image": "D.jpg" }
      ]
    },
    "exo_to_ego": { "…": "same shape, mirrored direction" }
  }
}

Note that Subtask 2's question_dir / option_dir are group-relative and already include the direction prefix (e.g. ego_to_exo/options), so resolve them directly against the group folder. The top-level question field, where present, is a group-level description rather than a model prompt; use sub_questions.<direction>.question as the prompt.

answers.json — ground truth

One file per subtask × scene, keyed by group id.

Subtasks 1, 3, 4:

{
  "group0000": {
    "correct_answer": "A",
    "correct_image": "A.jpg",
    "correct_frame_time_sec": 8.96
  }
}

Subtask 3 (indoor) additionally records correct_source_time_sec.

Subtask 2 nests one answer per direction:

{
  "group0000": {
    "sub_answers": {
      "ego_to_exo": { "correct_answer": "C", "correct_frame_time_sec": 29.1 },
      "exo_to_ego": { "correct_answer": "D", "correct_frame_time_sec": 29.1 }
    }
  }
}

These files contain only answer fields — no source-video names, clip paths, action labels, or other provenance that could reveal the target state. The image trees themselves are free of answer-bearing metadata: no per-group file states the correct option, and candidate entries carry no temporal or ordering cues.

Loading the data

import json
from pathlib import Path
from PIL import Image

root = Path("Subtask1_Indoor")
answers = json.loads((root / "answers.json").read_text(encoding="utf-8"))

group = root / "group0000"
meta = json.loads((group / "input_meta.json").read_text(encoding="utf-8"))

prompt = meta["question"]
frames = [Image.open(group / meta["question_dir"] / n).convert("RGB")
          for n in meta["question_images"]]
options = {o["label"]: Image.open(group / meta["option_dir"] / o["image"]).convert("RGB")
           for o in meta["options"]}

correct = answers["group0000"]["correct_answer"]      # "A"

For Subtask 2, iterate meta["sub_questions"] and resolve each direction's own question_dir / option_dir against the same group folder.

Recommended practice for the observation frames: keep them in the given temporal order and present all of them — temporal context and ordering are part of the task.

Licensing

The benchmark images are derived from six publicly available egocentric and ego–exo video sources, each with its own license and terms of use. This release inherits those terms: please consult and comply with the license of each source dataset before using, redistributing, or publishing results based on this benchmark. The questions, annotations, and metadata produced for EgoExo-Next are released for non-commercial research use.