File size: 3,834 Bytes
974e696
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
---
license: other
language:
- en
pretty_name: Event-Guided Video Depth Estimation Workshop Dataset
tags:
- computer-vision
- depth-estimation
- event-camera
- video
- low-light
task_categories:
- depth-estimation
source_datasets:
- DVD sde_in_release
- DVD sde_out_release
---

# Event-Guided Video Depth Estimation Workshop Dataset

This dataset is a mirrored and aligned workshop-ready version of the DVD event-guided video depth estimation data.

It packages each scene into a canonical folder tree that aligns:

- low-light RGB frames
- per-frame event slices
- a scene-level `lowlight_event.npz`
- the matched depth ground truth copied from `inference_results/*/normal/depth.npz`

The dataset is designed for direct upload to Hugging Face as a dataset repository.

The official competition split is scene-level and uses a 6:2:2 train/val/test ratio by video count. The split is encoded directly in the top-level `train/`, `val/`, and `test/` folders.

## Dataset Goals

This dataset is intended for event-guided video depth estimation in low-light conditions.

The expected input is:

- low-light event data
- low-light RGB images

The expected output is:

- aligned depth supervision for the corresponding scene

## Directory Layout

The canonical structure is:

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

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

Each scene contains:

```text
scene_name/
  low/
    <timestamp>.png
    <timestamp>.npz
    lowlight_event.npz
  normal/
    <timestamp>.png
    depth.npz
    manifest.json
```

## File Semantics

### Low-light RGB frames

The PNG files under `low/` are the low-light image sequence used as model input.

### Event slices

Each per-frame `.npz` file in `low/` stores a slice of the event stream. The slices are timestamp-aligned and may overlap with neighboring slices in the raw source data. During materialization, the overlap is trimmed at timestamp boundaries so that adjacent slices do not double-count the same events.

`lowlight_event.npz` stores the concatenated trimmed event stream for the scene.

### Depth ground truth

`normal/depth.npz` is copied from the corresponding inference result directory:

`inference_results/<release>_png_depth/<scene>/normal/depth.npz`

This file is used as the aligned depth target for evaluation and training.

## Data Preparation

The repository includes a preprocessing script:

```bash
python utils/process_workshop_data.py
```

By default it reads from:

- `data/sde_in_release`
- `data/sde_out_release`
- `inference_results`

and materializes the final dataset under:

- `../workshop_data`

## Intended Use

This dataset is meant for research on event-guided video depth estimation under low-light conditions.

Recommended usage:

- feed low-light RGB sequences together with event information as input
- use the copied `normal/depth.npz` as supervision or evaluation ground truth
- keep the scene-level temporal alignment intact when training or evaluating temporal models

## Notes

- The dataset tree is already aligned for scene-level consumption.
- The raw source event slices contain overlap; use the trimmed workshop copy instead of the raw source tree.
- The competition split is by scene/video count, not by frame count.
- The folder name itself is the split assignment; no separate top-level split manifest is required.

## Competition Draft

See [COMPETITION.md](COMPETITION.md) for a Codabench-style challenge description draft that follows the same page-tab structure as the reference competition.

## Citation

If you use this dataset, please cite the original DVD project and mention that you are using the workshop-aligned export.