OSResight/tartanair-episodes-v3-parquet

---
dataset_info:
features:
- name: episode_id
dtype: int32
- name: frame_idx
dtype: int32
- name: jpeg
dtype: binary
- name: depth
sequence: float32
- name: seg
sequence: uint16
- name: pose
sequence:
dtype: float64
length: 16
- name: intrinsics
sequence:
dtype: float32
length: 4
- name: gravity
sequence:
dtype: float32
length: 3
- name: scene
dtype: string
- name: trajectory
dtype: string
- name: dataset
dtype: string
- name: camera_model
dtype: string
- name: image_h
dtype: int32
- name: image_w
dtype: int32
- name: num_frames_in_episode
dtype: int32
config_name: default
splits:
- name: train
num_examples: 72000
license: bsd-3-clause
task_categories:
- robotics
tags:
- slam
- visual-odometry
- episode-data
- parquet
---

# tartanair-episodes-v3-parquet

Per-frame Parquet dataset for ReCAST tracker training.

## Schema

One row per frame, grouped by `episode_id`. Arrow memory-mapped access
enables reading specific frames without loading entire episodes.

| Column | Type | Description |
|--------|------|-------------|
| episode_id | int32 | Episode identifier |
| frame_idx | int32 | Frame index within episode |
| jpeg | binary | JPEG-encoded RGB frame |
| depth | list\<float32\> | Flat H×W depth map |
| seg | list\<uint16\> | Semantic segmentation (empty if absent) |
| pose | fixed\_size\_list\<float64, 16\> | 4×4 camera-to-world pose (vision convention) |
| intrinsics | fixed\_size\_list\<float32, 4\> | [fx, fy, cx, cy] in pixels |
| gravity | fixed\_size\_list\<float32, 3\> | World-frame gravity direction |
| camera_model | string | "pinhole" or "equidistant" |

## Conventions

- **Poses**: Vision convention (X=right, Y=down, Z=forward). Normalized during conversion.
- **Depth**: Projective z-depth (all datasets, normalized during conversion).
- **Gravity**: Unit vector pointing downward in world frame.

## Provenance

- Dataset: tartanair
- Episodes: 60
- Frames: 72000
- Created: 2026-03-15T14:19:10.370555+00:00
- Converter: `scripts/convert_episodes_to_parquet.py`