🛩️ Boread

A Dataset and Benchmark for Dynamic UAV Viewpoint Scheduling in Aerial-Ground Cooperative Perception

🤗 Hugging Face dataset · 🔗 Anonymous code and documentation

✨ CARLA-Air-based · 🛰️ Aerial-ground cooperative · 👁️ Visibility-aware · 🧭 Motion-constrained

🎯 Overview

Boread is a CARLA-Air-based dataset and benchmark for active viewpoint selection in aerial-ground cooperative perception. A UAV selects successive viewpoints to reveal targets occluded from a ground vehicle and improve cooperative 3D detection and tracking. Candidate aerial views observe the same evolving traffic scene, enabling controlled evaluation of sequential scheduling under shared motion constraints.

This release provides representative samples from the Boread dataset, including two scenes (GW0000 and GW0007) with complete temporal sequences across three UAV heights and eight candidate viewpoints per height.

✨ Key Features

• 🎬 Multi-view sequential data. 480 scene clips, 57.6K cooperative frames and approximately 300K images across eight candidate UAV viewpoints and flight heights of 20, 30 and 40 m.
• 📷 Synchronized aerial-ground sensing. Five UAV cameras complement four ground cameras and LiDAR, with aligned observations, calibration and 3D annotations at 10 Hz.
• 👁️ Viewpoint-conditioned visibility. Persistent instance IDs and per-view visibility annotations connect viewpoint selection to target observability and occlusion recovery.
• 🧭 Motion-constrained scheduling. Three decisions per 12-second episode couple perception utility with reachability, transition time and a 44 m movement budget.
• 🌍 Diverse traffic and occlusion. Urban, highway and rural scenes cover crossing, merging, large-vehicle occlusion and road-structure effects.
• 📊 Unified evaluation. Nine baselines span static, random, heuristic, planning and learned policies, with detection, continuous tracking, occlusion recovery and travel cost metrics.

📊 Dataset summary

Property	Release
Scenes	2: GW0000 and GW0007
UAV heights	20, 30 and 40 m
Candidate viewpoints	8 per height
Frames per sequence	120 at 10 Hz
UAV candidate sequences	48
RGB images	29,760
Image resolution	1920 × 1080
Vehicle LiDAR scans	240

Each ground frame contains four surround-view RGB cameras and one LiDAR scan. Each UAV frame contains front, back, left, right and bottom RGB cameras. Ground observations are shared across candidate UAV sequences for the same scene.

🌍 Released scenes

Scene	Map	Traffic setting
GW0000	Town10HD	Urban intersection with crossing traffic and building-corner occlusion
GW0007	Town04	Highway curve and merge with road-structure and moving-vehicle occlusion

📡 Sensors & annotations

The release contains synchronized RGB images, vehicle LiDAR, 3D bounding boxes, semantic categories, persistent instance identities, visibility measurements, camera calibration, poses and frame-level aerial-ground transforms.

Vehicle LiDAR is stored as little-endian float32 records in the order (x, y, z, intensity). Each 3D box label contains:

type x y z length width height roll pitch yaw track_id visibility

Boxes use a right-handed forward-left-up frame aligned with the corresponding agent pose. Cooperative boxes use the vehicle frame. Pixel-level instance segmentation masks are not included.

Scene manifests, height manifests and bridge records provide the released sequence structure and pair ground observations with UAV candidates by scene, height, viewpoint and frame index.

🧭 Benchmark scope

Boread supports sequential cooperative detection and tracking with temporal state carried across viewpoint changes. A scheduling policy selects feasible UAV viewpoints under a shared movement constraint.

The benchmark reports:

• 🎯 Average Precision (AP) for 3D detection
• 📈 Average Multi-Object Tracking Accuracy (AMOTA)
• 👁️ Occlusion Recovery Rate (ORR)
• 🧭 Mean UAV travel distance

ORR measures recovery of targets missed by the ground view but visible from at least one candidate UAV viewpoint.

💻 Code & documentation

Benchmark implementations, evaluation protocols and additional documentation are available at the anonymous repository.