README.md

---
license: cc-by-4.0
size_categories:
- 10K<n<100K
task_categories:
- video-classification
- video-text-to-text
tags:
- video
- features
- emotion
- event-detection
- nature
- eccv
---

# TRINITY Dataset

**Paper:** [TRINITY: A Multi-Perspective Benchmark for Personal-Style Video Highlight Detection](https://huggingface.co/papers/2608.29577)  
**Code:** [GitHub Repository](https://github.com/vanilladucky/TRINITY)

TRINITY accompanies our ECCV 2026 paper *"TRINITY: A Multi-Perspective Benchmark for Personal-Style Video Highlight Detection"* and provides precomputed video features for three complementary highlight-detection perspectives: **Emotion**, **Event**, and **Nature**.

Traditional highlight detection assumes a single, event-centric notion of saliency, which fails to generalize to unconstrained personal videos where highlights are heterogeneous and perspective-dependent. TRINITY decomposes highlight saliency into three complementary dimensions within a unified temporal framework:

- **Event** — event-driven semantic peaks, adopting the replay-based annotations from Mr. HiSum.
- **Emotion** — affective facial moments, annotated via a scalable two-stage automatic pipeline (frame-level facial expression recognition + multimodal LLM verification).
- **Nature** — aesthetic scenic highlights, annotated via scenic localization (person/landscape detection) followed by frame-level aesthetic scoring. Features are sampled at 1 FPS, same as Event.

The three perspectives are intentionally complementary rather than redundant (cross-perspective IoU as low as 0.002–0.052 on Mr. HiSum), enabling structured analysis of heterogeneous highlight patterns beyond the traditional single-perspective paradigm. This repository provides only the precomputed CLIP ViT-B/32 features (and Emotion ground-truth labels) used in the paper's experiments — see the paper for full annotation pipeline details.

## Subsets

| Subset  | Feature granularity        | Feature dim | Dtype   | Ground truth |
|---------|-----------------------------|-------------|---------|--------------|
| Emotion | 5-second averaged windows  | 512         | float16 | binary highlight labels + segment metadata (separate archive) |
| Event   | 1 FPS                       | 512         | float32 | frame-level replay-saliency scores (separate archive) |
| Nature  | 1 FPS                       | 512         | float16 | frame-level aesthetic scores (separate archive); includes `_aug` augmented variants |

All features are stored per-clip as individual `.npz` files, packaged into numbered zip shards for download.

### Emotion

- `emotion/emotion_feature_{1..4}.zip` — per-clip feature files, each `.npz` with key `features` of shape `(T, 512)`.
- `emotion/emotion_gt.zip` — matching ground-truth files, each `.npz` containing:
  - `labels`: shape `(T,)`, float32
  - `clip_start`, `clip_end`, `clip_duration`: clip timing info
  - `num_segments`, `segment_duration`: segmentation info
  - `is_negative`: bool flag for negative samples
- `emotion/emotion_split.json` — `train_keys` / `val_keys` / `test_keys` listing the relative `.npz` filenames for each split (11,352 / 1,615 / 3,290 clips).

Feature and ground-truth files share the same filename across the two archive sets.

### Event

- `event/event_features_part_{1..145}.zip` — per-video feature files, each `.npz` with key `data` of shape `(T, 512)`, float32, sampled at 1 FPS.
- `event/event_gt_part_1.zip` — matching ground-truth files, each `.npz` with key `data` of shape `(T,)`, float64, giving the frame-level replay-based saliency score (one value per 1-second frame, aligned with the feature sequence).
- `event/mr_hisum_split_original.json` — `train_keys` / `val_keys` / `test_keys` (27,892 / 2,000 / 2,000 videos).

Feature and ground-truth files share the same filename (e.g. `video_1.npz`) across the two archive sets.

### Nature

- `nature/nature_feat_part_{1..32}.zip` — per-video feature files, each `.npz` with key `features` of shape `(T, 512)`, sampled at 1 FPS. Files suffixed `_aug` are augmented variants of their base video.
- `nature/nature_gt_part_1.zip` — matching ground-truth files, each `.json` containing a list of per-frame aesthetic saliency scores in `[0, 1]`, aligned with the feature sequence.
- `nature/nature_split.json` — `train_keys` / `val_keys` / `test_keys` (12,763 / 874 / 1,903 clips).

Feature and ground-truth files share the same base filename (e.g. `--rd0X_LQCo.npz` / `--rd0X_LQCo.json`) across the two archive sets.

## Usage

Download and unzip the shards for the subset(s) you need:

```bash
hf download vanilladucky/TRINITY --repo-type dataset --local-dir ./TRINITY
cd TRINITY/nature && for f in *.zip; do unzip -q "$f" -d nature_feat; done
```

Load a feature file:

```python
import numpy as np
feat = np.load("nature_feat/--rd0X_LQCo.npz")["features"]  # (T, 512)
```

## Citation

If you use this dataset, please cite our paper:

```bibtex
@inproceedings{chen2026trinity,
  title     = {{TRINITY}: A Multi-Perspective Benchmark for Personal-Style Video Highlight Detection},
  author    = {Chen Qianqian and Kim Hyun Bin and Wijaya Denzel Elden and Yi Yang and Liu Bo and Ding Yangkai},
  booktitle = {European Conference on Computer Vision (ECCV)},
  year      = {2026}
}
```

## License

Released under CC BY 4.0. Underlying source videos belong to their original datasets/owners (e.g. Mr. HiSum, YouTube); this release distributes only extracted features and derived annotations.