← Paper Digest

AGE-Capture: Autonomous, General, and Efficient UAV Capture for High-Fidelity Scene Reconstruction

Xiaofeng Jin, Ningbo Bu, Matteo Frosi, Yiran Guo, Weilun Ren, Matteo Matteucci

November 2026

TL;DR AGE-Capture flies a UAV to collect images for Gaussian-splatting reconstruction. It picks views with a lightweight renderability state and stops by itself once further views add little. In ten simulated scenes it beats two online baselines by 3.30-3.53 dB PSNR while using 58.2% of the 900 s budget, and it runs onboard a real quadrotor in nine scenes.

What they do

Most active-reconstruction methods decide where to look next, then stop at a fixed view count or time budget. AGE-Capture builds a complete UAV capture loop around R3CON’s lightweight renderability state. The loop scores candidate views, decides when a region and finally the whole mission has nothing left worth capturing, and keeps planning while the drone flies. It accepts pinhole, fisheye or panoramic RGB cameras, with depth maps or LiDAR for geometry.

How it works

  • Shared ray interface. Every camera model is converted to world-space rays. A tile-based GPU ray tracer finds the visible surface for each ray, and a per-camera field-of-view footprint on a spherical direction grid scores candidate orientations.
  • Regions and stopping. Reachable free space is split into seeded regions. A position whose 90th-percentile renderability gain exceeds a threshold is a reconstruction frontier. Regions without one are dropped, and the mission ends when the queue is empty.
  • View selection. Within a region, position-direction pairs are ranked by area-weighted renderability gain per unit of estimated flight and planning time.
  • Asynchronous capture. The next plan starts before the current trajectory ends. A temporary pseudo-update for the pending view keeps the planner from choosing the same surfaces again.

Results

  • In simulation (Habitat, ten scenes, 900 s cap) mean PSNR is 25.53 dB, against 22.00 for ActiveGS and 22.23 for TRACE. Missions end after 152.8-837.2 s, and mean peak GPU memory is 1.44 GiB versus 15.93 and 21.61 GiB.
  • A panoramic camera instead of a pinhole adds 3.08 dB and cuts mission time from 559.2 s to 229.8 s.
  • Without early replanning, missions take 40.7% longer. Without the pseudo-update they record 27.3% more observations.
  • On a real quadrotor (onboard Jetson Orin NX) in nine indoor and outdoor scenes, the loop captured 49-115 frames per scene. PSNR on 100 held-out images per scene was 18.30-28.96 dB.

Limitations

The real flights are not compared with any baseline, and their splats are trained offline from recorded data. To fit one flight, the capture region was limited and the stopping threshold relaxed, which left distant surfaces under-observed. Vegetation, reflections and view-dependent appearance still show blur or artifacts.

Figures

From Figure 2: System overview. RGB rays from any camera model enter a shared observation interface. Global region ma...
From Figure 2: System overview. RGB rays from any camera model enter a shared observation interface. Global region management, regional next-best-view planning and event-driven capture use the renderability state to select and schedule viewpoints, while the exploration and trajectory modules provide geometry and motion.
From Figure 5: Global region management. The upper panels show the free-space distance field, the seed-centered regio...
From Figure 5: Global region management. The upper panels show the free-space distance field, the seed-centered region partition and the initial visit order. Below, completed regions are check-marked, and a region that fails the reconstruction-frontier test is removed before the remaining regions are reordered from the UAV position.
From Figure 8: Online reconstructions under the common 900 s mission-time cap in six scenes, comparing ground truth, ...
From Figure 8: Online reconstructions under the common 900 s mission-time cap in six scenes, comparing ground truth, ActiveGS, TRACE and AGE-Capture. White boxes mark baseline regions with missing geometry or appearance degradation.
From Figure 15: Nine real-world test environments. Each panel combines a LiDAR point-cloud map, the UAV trajectory an...
From Figure 15: Nine real-world test environments. Each panel combines a LiDAR point-cloud map, the UAV trajectory and the capture positions; insets show onboard RGB observations.
Abstract

High-fidelity scene reconstruction with an unmanned aerial vehicle (UAV) requires more than identifying valuable viewpoints: the system must autonomously acquire informative observations and determine task completion under different sensing configurations and limited flight resources. We present AGE-Capture, an autonomous, sensor-adaptive, and efficient active-capture framework for real UAVs. Building on renderability, the framework uses a lightweight online state to evaluate the utility of candidate views, assess task progress, and guide flight execution. By decoupling geometric exploration from reconstruction observations and establishing a unified camera-ray interface, the framework allows different range sensors and RGB camera models to share a reconstruction-driven acquisition pipeline. Reconstruction-frontier viewpoints characterize residual acquisition value for region screening and autonomous termination; dynamics-aware hierarchical planning combined with asynchronous event-driven capture reduces unnecessary inter-region motion, time spent waiting for plans, and repeated observations. Across ten indoor and outdoor simulation scenes, AGE-Capture achieves a mean peak signal-to-noise ratio (PSNR) $3.30$--$3.53\,\mathrm{dB}$ higher than that of the online baselines while using $58.2\%$ of the common mission-time budget on average. Its mean peak GPU memory usage is $91.0\%$--$93.3\%$ lower than that of the same baselines. Experiments with different sensor configurations validate the adaptability of the shared acquisition pipeline, while flights in nine real-world scenes further demonstrate onboard closed-loop operation under practical computational and motion constraints.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.