---
pretty_name: LayeredFlow-Syn Extracted Ground Truth
license: other
size_categories:
- n<1K
tags:
- synthetic
- ground-truth
- optical-flow
- layered-scene
configs:
- config_name: default
  data_files:
  - split: train
    path: data/**/*.parquet
dataset_info:
  features:
  - name: visualization
    dtype: image
  - name: sample_id
    dtype: string
  - name: scene_id
    dtype: string
  - name: relative_path
    dtype: string
  - name: num_bytes
    dtype: int64
  - name: sha256
    dtype: string
  - name: content
    dtype: binary
  splits:
  - name: train
    num_examples: 15549
---

# LayeredFlow-Syn Extracted Ground Truth

This repository contains extracted LayeredFlow ground-truth annotations in Parquet format.

## Layout

```text
data/<scene>/<sample>.parquet
```

For example:

```text
data/0/0_0.parquet
```

Each Parquet file stores one extracted sample. Rows correspond to files from the extracted
sample directory and include a leftmost `visualization` image preview when available,
`relative_path`, `num_bytes`, `sha256`, and binary `content`.
Rows are ordered by `frame0/left`, `frame0/right`, `frame1/left`, then `frame1/right`.
Inside each frame/view group, flow visualization PNG rows named `*_visualization.png`
come first. `.flo` rows also use the matching `*_visualization.png` as their
leftmost preview image, while mask/position PNG rows preview themselves.

Converted selector `20` contains `515` samples.