COMPETITION.md

# Event-Guided Video Depth Estimation Competition

## Overview

This challenge focuses on event-guided video depth estimation in low-light scenes.

Participants are asked to use low-light RGB frames together with event streams to predict dense depth maps for each scene. The benchmark is organized at the scene/video level so that temporal consistency is preserved across frames.

The official split follows a 6:2:2 train/val/test ratio by video count. The split assignment is encoded directly in the top-level directory structure.

## Dataset

The competition dataset is a workshop-aligned mirror of the DVD event-guided depth data.

Each scene contains aligned RGB frames, event slices, and depth supervision in the following layout:

```text
train/
  <scene_name>/
    low/
    normal/
val/
  <scene_name>/
    low/
    normal/
test/
  <scene_name>/
    low/
    normal/
```

Each scene directory includes a `manifest.json` file with the release and split metadata.

Training and validation scenes are intended for development and public benchmarking. Test labels should remain hidden during the competition phase.

## Submission

Participants should submit one predicted depth sequence for each test scene.

Recommended submission layout:

```text
submission/
  <scene_name>/
    normal/
      depth.npz
```

Each `depth.npz` file should contain a depth array aligned with the scene frame order. A 4D tensor shaped `[T, H, W, 1]` is preferred for compatibility with the existing DVD tooling, although `[T, H, W]` can also be supported by the evaluation script.

## Evaluation

The current DVD evaluation code reports the following depth metrics:

- AbsRel / `abs_relative_difference`
- RMSE / `rmse_linear`
- delta1 / `delta1_acc`
- delta2 / `delta2_acc`
- delta3 / `delta3_acc`
- SILog / `silog_rmse`

For launch, we suggest selecting one primary ranking metric from the list above and reporting the rest as diagnostics.

## Rules & Timeline

- Only use the training data and any additional resources explicitly allowed by the organizers.
- Do not inspect or annotate hidden test labels.
- Keep scene-level temporal ordering intact when producing predictions.
- Submission quota: TBD.
- Development phase: TBD.
- Final test phase: TBD.

## Terms

- Dataset license: TBD.
- Code and baseline license: TBD.
- By participating, teams agree to the competition rules and the platform terms.
- The organizers may update the submission format or evaluation script with prior notice.