---
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.parquet
  - split: test
    path: data/test.parquet
task_categories:
- object-detection
---

# eAP Detection Benchmark Dataset

This public dataset repo contains eAP detection inputs for the released train/test benchmark splits.

## Files

- `data/train.parquet`: public `train` split metadata.
- `data/test.parquet`: public `test` split metadata.
- `data/<split>/<sequence_id>/events.h5`: event stream for the sequence.
- `data/<split>/<sequence_id>/rgb_shards/rgb-00000.tar`: RGB PNG tar shards.
- `data/train/<sequence_id>/labels.parquet`: public train labels.
- `sample_submission.json`: eAP-native submission template.

Test labels are not included in this public dataset. Evaluation is handled by the CodaBench challenge.

## Parquet Columns

`sample_token`, `split`, `sequence_id`, `rgb_shard_path`, `rgb_member_path`, `events_path`, `labels_path`,
`rgb_exposure_start_timestamp_us`, `rgb_exposure_end_timestamp_us`, `K_event`, `T_event_ego`.

`rgb_shard_path` is relative to the repository root. `rgb_member_path` is the member path inside that tar.
Tar members use `rgb/<image_name>.png`; events remain as one `events.h5` file per sequence and are not tarred.
For test rows, `labels_path` is null by design.

## Loading Example

```python
from datasets import load_dataset

ds = load_dataset("parquet", data_files={"train": "data/train.parquet", "test": "data/test.parquet"})
```

## Submission Format

Submit a JSON object with `meta` and `results`. `results` is keyed by `sample_token`; each detection uses:
`class_name`, `translation_m` (`[x, y, z]`), `size_lwh_m` (`[length, width, height]`), `yaw_rad`,
`velocity_mps` (`[vx, vy]`), and `score`.

## Splits

- `train`: 118247 samples from 40 sequences.
- `test`: 35940 samples from 12 sequences.