Synthetic drone data cuts costs and boosts aerial detection accuracy
Dreams Are Cheaper Than Flights: Curated World-Model Data Improves Aerial Detection
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Collecting real drone video data to teach computers to spot animals or objects can be slow, costly, and disruptive in nature. The authors used a very large AI model to generate new, synthetic drone videos based on real footage. Training detectors with this extra synthetic data improved their accuracy more than using real data alone, and it was cheaper than collecting more real videos. Even though many generated clips were rejected by experts, the cost saved on fieldwork was substantial. This shows that creating fake but realistic video data could help train better aerial detection systems.
What this means in practice
- •For wildlife monitoring teams: Reduce expensive field data collection by supplementing drone video datasets with AI-generated synthetic footage for better animal detection accuracy.
- •For urban surveillance operators: Improve object detection in city drone footage by training with curated synthetic videos modeled on existing real data.
Authors
Rugved Katole, Vedant Patil, Jenna Kline, Christopher Stewart
Abstract
Robot perception is increasingly data driven, yet field data is slow, costly, and sometimes disruptive to collect. Video world models offer a cheaper source: seeded with real frames, they extend a recorded scene into new footage with the deployment's sensor, altitude, and environment. We ask whether such synthetic data matches the utility of real field data and whether detectors trained with it perform in the field. Seeding a 7B-parameter world model with wildlife drone video, we more than double the training set with curated frames. On held-out videos from the same drones, mAP@50--95 rises 4.3\% when these frames are added to real data and 10.3\% when they replace it, and up to 33.5\% on an urban benchmark. Although an expert rejects an estimated 87.5\% of generated clips, the data costs about \$1,100 of GPU time, against more than \$3,200 of field labor alone for a less useful real training set. Synthetic data thus shifts data costs from fieldwork to compute.