File size: 1,639 Bytes
ecdac4d
 
 
49f7c2d
ecdac4d
40ecd6c
ecdac4d
 
49f7c2d
ecdac4d
 
 
 
 
 
49f7c2d
ecdac4d
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
---
license: mit
---
# EgoWalk Trajectories

This is the extracted version of the EgoWalk dataset, obtained from the [raw one](https://huggingface.co/datasets/EgoWalk/raw-recordings).
Please check the [API for this dataset](https://github.com/egowalk/egowalk-dataset) for the basic usage.

## Data format
The data is mainly organized in following directories:
- *data*: includes csv with the main per-frame annotations - video name, frame index, UTC timestamp, camera pose (may be `null` due to odometry failures, see next sections)
- *video*: includes video files that are referenced in *data*. Lossy encoding is used for RGB channel, lossless - is for Depth channel.
- *annotations*: includes language goal annotations for the selected frames. We release to versions of annotations: *normal* - the main version, and *brief* - the version with slightly more brief sentences.
- *meta*: includes additional metadata, namely, approximate camera heights, camera params, list of trajectories, etc.

## Notes on the odometry
[ZED SDK](https://www.stereolabs.com/en-ru/developers/release) was used to calculate the odometry from the raw data. This odometry is still imperfect, and sometimes may result in odometry failures of two kinds:
- `nulls`: odometry method failed to compute the pose at this particular timestamp
- Reinitialization: due two large amount of sequential failures odometry was re-initialized in the middle of trajectory, and re-initialized pose starts in zero.
API includes some mechanisms to process such cases, such as [cutters](https://github.com/egowalk/egowalk-dataset/blob/main/egowalk_dataset/datasets/gnm/cutters.py).