Research map / Detection and segmentation / Open-set and few-shot detection
Multiview Aerial Visual RECognition (MAVREC): Can Multi-view Improve Aerial Visual Perception?
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces MAVREC, a synchronized ground- and aerial-view video dataset from European landscapes, and benchmarks supervised and curriculum-based semi-supervised object detectors, finding ground-view pre-training helps aerial detection.
Problem
Aerial object detection from UAVs performs poorly on existing datasets, which are small, low-resolution, and geographically biased toward Asia and North America; models trained on ground-view data also generalize poorly to aerial views.
Approach
Record time-synchronized ground and drone videos in European geographies and show that pre-training on same-geography ground-view images, combined with a curriculum-ordered teacher-student semi-supervised detector, improves aerial detection.
Key results
- Fine-tuning D-DETR on MAVREC aerial view after MAVREC ground-view pre-training reaches 44.8 AP on the test set, versus 33.2 with MS-COCO pre-training and 20.9 (validation) with VisDrone.
- The curriculum-based semi-supervised D-DETR improves over the semi-supervised baseline by 8.5% (validation) and 7% (test) on aerial view.
- MAVREC contains about 537K frames, 1.1M bounding boxes, and 2.7K resolution video across 11 European locations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Pre-train D-DETR on MAVREC ground-view images before aerial fine-tuning rather than MS-COCO | D-DETR · MAVREC aerial test | AP: 33.2 -> 44.8 | Helped |
| Use curriculum training with labeled ground images before labeled aerial images | Omni-DETR · MAVREC aerial validation | AP: 29.3 -> 37.8 | Helped |
| Use curriculum training with labeled ground images before labeled aerial images | Omni-DETR · MAVREC aerial test | AP: 19.8 -> 23.2 | Helped |
| Add unlabeled ground images during the second training phase (ablation) | Curriculum learning with Omni-DETR · MAVREC aerial test | AP: 23.2 -> 26.7 | Helped |
| Use curriculum training with labeled ground images before labeled aerial images | Omni-DETR · MAVREC aerial validation | AP50: 49.3 -> 64.9 | Helped |
Surprising findings
- Grounding-DINO, pretrained on large corpora, underperforms on MAVREC aerial images (20.4 AP test) relative to ground-view-pretrained D-DETR.
- Yolo-NAS, which outperforms other YOLO models on ground images, underperforms YOLOv7 on aerial images.
Limitations
- Data collected in only 11 European locations during spring and summer, limiting geographic diversity.
- Annotation error around 6% estimated from a review sample of 600 images.
- Ground view and aerial view are not perfectly time-synchronized due to manual alignment.
Experimental setup
Datasets: MAVREC MS-COCO VisDrone
Benchmarks: MAVREC
Metrics: mAP AP50 APS APM
Baselines: DETR Deformable DETR YOLOv7 Yolo-NAS Grounding-DINO VisDrone pre-training MS-COCO pre-training
Builds on: D-DETR Omni-DETR Grounding DINO
Closest papers by meaning
- AGC-Drive: A Large-Scale Dataset for Real-World Aerial-Ground Collaboration in Driving Scenarios (NeurIPS 2025 · similarity 0.86)
- RemDet: Rethinking Efficient Model Design for UAV Object Detection (AAAI 2025 · similarity 0.86)
- Unleashing Multispectral Video's Potential in Semantic Segmentation: A Semi-supervised Viewpoint and New UAV-View Benchmark (NeurIPS 2024 · similarity 0.86)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · similarity 0.85)
- FBRT-YOLO: Faster and Better for Real-Time Aerial Image Detection (AAAI 2025 · similarity 0.85)
- Zero-Shot Aerial Object Detection with Visual Description Regularization (AAAI 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.