Scout Dataset Toolkit

A comprehensive Python toolkit for working with the Scout multi-camera pedestrian tracking dataset. This toolkit provides easy-to-use interfaces for loading annotations in multiple formats, visualizing data, and evaluating tracking methods.

📦Download the dataset

The Scout dataset is a large-scale multi-camera pedestrian tracking dataset captured in realistic outdoor environments. It can be downloaded from https://scout.epfl.ch/download.

This toolkit provides:

• Multi-format data loading - Support for COCO, Individual, and MOT annotation formats
• Rich visualization tools - 2D bounding box visualization and 3D world coordinate plotting
• Comprehensive evaluation metrics - Both 3D world coordinate and standard MOT metrics
• Camera calibration utilities - 3D geometry operations and camera projections
• PyTorch integration - Ready-to-use PyTorch datasets and data loaders

🔧 Installation

Option 1: Conda Environment (Recommended)

conda env create -f environment.yml
conda activate scout-env

Then install trackeval:

pip install --no-deps git+https://github.com/JonathonLuiten/TrackEval.git@12c8791b303e0a0b50f753af204249e622d0281a

Option 2: pip Installation

pip install -r requirements.txt

🚀 Quick Start

Check out example.ipynb for a comprehensive walkthrough of the toolkit's main features:

from src.loading import Individual
from src.visualisation import draw_annotation

# Load annotations for a sequence
loader = Individual(root_dir=".", sequence=1)
annotations = loader.retrieve(frame_id=81, target_cam_id='cam_2')

# Visualize annotated frame
img = loader.annotate_frame(81, 'cam_2')
loader.show_frame(img)

📁 Key Files and Modules

🔄 Data Loading (src/loading.py)

Support for multiple annotation formats:

• Coco - COCO format annotations with full sequence loading
• Individual - Frame-by-frame text file annotations with caching
• MOT - MOT Challenge format for tracking evaluation
• AnnotationDataset - PyTorch-compatible dataset wrapper

# Load different annotation formats
from src.loading import Coco, Individual, MOT

# COCO format - full sequence
coco = Coco(rootdir, sequence=1)
annotations = coco.retrieve(frame_id=0, target_cam_id='cam_0')

# Individual format - per-frame files
individual = Individual(rootdir, sequence=1)
annotations = individual.retrieve(frame_id=81, target_cam_id='cam_2')

# MOT format - tracking sequences
mot = MOT(rootdir, sequence=1)
annotations = mot.retrieve(frame_id=0, target_cam_id='cam_0')

📊 Visualization (src/visualisation.py)

Comprehensive visualization tools:

• draw_annotation() - Draw 2D bounding boxes with track IDs
• draw_3d() - Plot 3D world coordinates with mesh overlay
• get_color_for_instance() - Consistent color mapping for track IDs

from src.visualisation import draw_annotation, draw_3d
import matplotlib.pyplot as plt

# 2D visualization
image = draw_annotation(image, bbox, track_id)
plt.imshow(image)
plt.show()

# 3D visualization with mesh
draw_3d(world_points, mesh)

📈 Evaluation Metrics (src/metric.py)

Comprehensive evaluation suite:

• MOTMetricEvaluator - Standard MOT metrics (MOTA, MOTP, IDF1, HOTA)
• compute_mot_metric() - 3D world coordinate evaluation
• scout_to_dataframe() - Convert Scout annotations to evaluation format
• simulate_predictions() - Generate synthetic predictions for testing

from src.metric import MOTMetricEvaluator, compute_mot_metric

# MOT evaluation
evaluator = MOTMetricEvaluator()
metrics = evaluator.compute_metrics("SCOUT")

# 3D world coordinate evaluation  
gt_df = scout_to_dataframe(gt_sequence, use_world_coords=True)
pred_df = scout_to_dataframe(pred_sequence, use_world_coords=True)
metrics = compute_mot_metric(gt_df, pred_df, threshold=1.0, nb_gt=len(gt_df))

🔢 Geometry Operations (src/geometry.py)

3D geometry and camera calibration utilities:

• project_world_to_camera() - Project 3D world points to 2D image coordinates
• project_2d_points_to_mesh() - Project 2D points to 3D mesh surface
• bbox_to_mesh() - Convert 2D bounding boxes to 3D world coordinates
• load_calibration() - Load camera calibration parameters

from src.geometry import project_world_to_camera, bbox_to_mesh
from src.loading import load_calibration

# Load camera calibration
calibration = load_calibration('.', 'cam_0', sequence=1)

# Project 3D world point to 2D image
world_point = [28.07, 180.42, -1.27]
pixel_coords = project_world_to_camera(world_point, calibration)

# Project 2D bounding box to 3D mesh
bbox = {'x': 807.0, 'y': 359.0, 'w': 13.0, 'h': 31.0}
world_coords = bbox_to_mesh(bbox, calibration, mesh)

🔥 PyTorch Integration (src/pytorch_dataset.py)

Ready-to-use PyTorch datasets:

from src.loading import AnnotationDataset, collate_fn
from torch.utils.data import DataLoader

# Create PyTorch dataset
dataset = AnnotationDataset(loader, frame_ids, cam_ids)
dataloader = DataLoader(dataset, batch_size=8, shuffle=True, 
                       collate_fn=collate_fn, num_workers=4)

# Iterate through batches
for batch in dataloader:
    images = batch['images']        # Tensor of shape (B, C, H, W)
    bboxes = batch['bboxes']        # List of bounding boxes
    world_coords = batch['world']   # List of world coordinates
    track_ids = batch['track_ids']  # List of track IDs

📖 Usage Examples

Basic Data Loading and Visualization

See example.ipynb for:

• Loading annotations in different formats
• Visualizing 2D annotations and 3D world coordinates
• Working with camera calibration
• PyTorch dataset integration

Evaluation and Metrics

See evaluation.ipynb for detailed examples of:

• 3D World Coordinate Evaluation: Using compute_mot_metric() for evaluation in 3D space
• Standard MOT Evaluation: Using MOTMetricEvaluator for MOTA, MOTP, IDF1, and HOTA metrics
• Synthetic Data Generation: Creating test predictions with simulate_predictions()
• Data Format Conversion: Converting between Scout format and evaluation formats

🎯 Dataset Information

Dataset Access

• Main Website: https://scout.epfl.ch/
• Research Group: EPFL Computer Vision Laboratory

Dataset Features

• Multi-camera setup: Synchronized cameras with overlapping fields of view
• 3D world coordinates: Precise 3D localization of pedestrians
• Multiple annotation formats: COCO, Individual, and MOT format support
• Camera calibration: High-precision camera calibration parameters
• Realistic scenarios: Outdoor pedestrian tracking in challenging conditions

Associated Research

• Paper: "One Graph to Track Them All: Dynamic GNNs for Single- and Multi-View Tracking"
• Arxiv: https://arxiv.org/abs/2507.08494
• Code Repository: https://github.com/cvlab-epfl/umpn

Citation

If you use this dataset in your research, please cite:

@article{engilberge25scout,
  title = {One Graph to Track Them All: Dynamic GNNs for Single- and Multi-View Tracking},
  author = {Martin Engilberge and Ivan Vrkic and Friedrich Wilke Grosche 
            and Julien Pilet and Engin Turetken and Pascal Fua},
  journal = {arXiv preprint arXiv:2507.08494},
  year = {2025}
}

📝 License

Please refer to the official Scout dataset website for licensing information and usage terms.