# SkyEvents

Multi-modal aerial dataset combining event camera and RGB recordings from a UAV (Anthony-ECPKN, ICLR 2026).

## Overview

| Field | Value |
|---|---|
| Sensors | Prophesee EVK4 (1280×720 events) + DJI RGB |
| Sessions claimed in paper | 45 |
| Sessions currently organized | 40 (27 in-paper + 13 extras) |
| In-paper trainable parent flights | 18 / 18 |
| Extras with complete event + RGB | 4 / 13 |
| Total size | ~2.0 TB |
| Integrity verification | One `SHA256SUMS.txt` per session |

A `session` is a single recording. A `flight` is one complete flight. The DJI Fly app auto-splits an RGB clip when it reaches ~17 GB, so one flight can map to one EVK4 event recording plus several RGB segments. 

## Layout

```
SkyEvents/
├── README.md                       # this file
├── manifest.csv                    # session-level index for all 40 entries
├── SHA256SUMS.INDEX.md             # how to use the SHA256 checksums
├── SkyEvents_actual_inventory.md   # actual session inventory (Chinese)
├── SkyEvents_download_summary.md   # original download sources (Chinese)
│
├── raw/                            # 18 parent flights from the 27 in-paper sessions
│   └── {YYYYMMDD_HHMM}/            # session_id prefers the event-raw timestamp (minute)
│       ├── events.raw              # Prophesee EVK4 continuous recording
│       ├── events.bias
│       ├── rgb.MP4    OR    rgb_part1.MP4 + rgb_part2.MP4
│       ├── rgb.LRF    OR    rgb_part1.LRF + rgb_part2.LRF
│       └── SHA256SUMS.txt          # SHA256 of every file in this session
│
└── extras/                         # 13 out-of-paper sessions (single-modality or short clips)
    └── {YYYYMMDD_HHMM}/            # includes one multi-segment RGB case (20251021_1611)
```

## Conventions

- `session_id = YYYYMMDD_HHMM`. Prefer the event-raw timestamp; fall back to the DJI RGB timestamp when no event is present.
- In `manifest.csv`, `kind=continuation` marks an RGB continuation; `continuation_of` points to its parent. Continuation RGB files are physically merged into the parent directory as `rgb_part1.*` and `rgb_part2.*`.
- Trainability is judged per parent: a parent needs `events.raw` and at least one RGB segment.
- This directory is a verbatim mirror of the upstream dataset and contains no detection labels or preprocessing artifacts.

## Integrity Verification

Every session directory ships with a `SHA256SUMS.txt` generated by GNU `sha256sum`. To verify one session:

```bash
cd raw/20250912_1549
sha256sum -c SHA256SUMS.txt
```

For full-set verification see [`SHA256SUMS.INDEX.md`](SHA256SUMS.INDEX.md).