GlobalTrace

GlobalTrace is a benchmark of GPS trajectories drawn from 73 cities/regions across the world, with masked-segment annotations for evaluating trajectory reconstruction methods. Three masking strategies are provided per trajectory: destination, large_gap, and simple_interpolation (200m gap target).

Splits

Split	Cities	Trajectories
train	44	8,815
validation	10	1,787
test	19	1,707

The validation/test splits are held out for evaluation; the test set is unchanged from the original release. See redistribution_summary.md for the dev-set redistribution details.

Schema (one row per trajectory)

Column	Type	Description
trajectory_id	string	{split}/{city}/{strategy}/{base_name}
split	string	train / validation / test
city	string	Source city or region
mask_strategy	string	destination, large_gap, or simple_interpolation
base_name	string	Original trace base name
activity_type	string	e.g. walking, hiking, driving, cycling
activity_classification_method	string	How the activity label was derived
mask_type	string	Same as mask_strategy (from metadata)
num_masked_segments	int32	Number of masked segments
num_masked_points	int32	Total points removed across all masked segments
original_gpx	string	Full original GPX (XML)
masked_gpx	string	GPX with the target segment(s) removed
metadata_json	string	JSON-encoded metadata: per-segment indices, start/end points, context_before/after, original_points

The full per-segment metadata is preserved in metadata_json (parse with json.loads).

Usage

from datasets import load_dataset

ds = load_dataset("joey234/globaltrace")
print(ds)
print(ds["validation"][0]["trajectory_id"], ds["validation"][0]["activity_type"])

import json
meta = json.loads(ds["validation"][0]["metadata_json"])
print(meta["masked_segments"][0]["start_point"])

The original per-city directory layout is also kept in this repo under train_masks_200m/, dev_masks_200m/, test_masks_200m/ for users who prefer to consume raw .gpx and metadata files directly.

Citation

If you use GlobalTrace, please cite the corresponding paper.