Research map / Detection and segmentation / CNN features
FBRT-YOLO: Faster and Better for Real-Time Aerial Image Detection
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.
Proposes FBRT-YOLO, a family of real-time detectors for aerial images that adds a Feature Complementary Mapping module (FCM) and Multi-Kernel Perception unit (MKP) on a slimmed YOLOv8 baseline, evaluated on VisDrone, UAVDT and AI-TOD.
Problem
Real-time object detection in high-resolution aerial images struggles with small targets and with balancing accuracy against the computational limits of embedded flight devices.
Approach
Mixes shallow spatial information into deeper backbone stages via a split-transform design with channel and spatial complementary mapping (FCM), replaces the final downsampling with a multi-kernel depthwise-plus-pointwise unit (MKP), and decouples downsampling with group convolution to remove redundancy.
Key results
- On VisDrone, FBRT-YOLO-X reaches 30.1 AP with 22.8M params and 185.8 GFLOPs, and FBRT-YOLO-N reaches 20.2 AP with 0.9M params at 192 FPS.
- FBRT-YOLO-S reduces parameters by 74% versus YOLOv8-S (2.9M vs 11.2M) while improving AP from 23.6 to 25.9 on VisDrone.
- On UAVDT, FBRT-YOLO reports 18.4 AP, above CEASC (17.1) and GLSAN (17.0).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace YOLOv8 with FBRT-YOLO-X | YOLOv8-X · VisDrone | AP: 28.9 -> 30.1 | Helped |
| replace YOLOv8-S with FBRT-YOLO-S | YOLOv8-S · AI-TOD | AP: 19.1 -> 20.2 | Helped |
| reduce inherent network redundancy (ablation) | YOLOv8-S · VisDrone | AP: 23.6 -> 23.4 | Hurt |
| add FCM after redundancy reduction (ablation) | YOLOv8-S with RR · VisDrone | AP: 23.4 -> 24.3 | Helped |
| add MKP to RR + FCM (ablation) | YOLOv8-S with RR + FCM · VisDrone | AP: 24.3 -> 25.9 | Helped |
| combine channel and spatial complementary mapping (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.0 -> 25.9 | Helped |
| use channel mapping only instead of no mapping (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.0 -> 25.6 | Helped |
| use spatial mapping only instead of no mapping (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.0 -> 25.3 | Helped |
| set FCM split ratios to (0.75,0.75,0.25,0.25) rather than (0.25,0.25,0.25,0.25) (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.5 -> 25.9 | Helped |
| use mixed MKP kernel sizes (3,5,7) rather than uniform (3,3,3) (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.4 -> 25.9 | Helped |
| use mixed MKP kernel sizes (3,5,7) rather than uniform (5,5,5) (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.6 -> 25.9 | Helped |
| use mixed MKP kernel sizes (3,5,7) rather than uniform (7,7,7) (ablation) | FBRT-YOLO-S · VisDrone | AP: 25.8 -> 25.9 | Helped |
Limitations
- Ablation and split-ratio experiments are run only on VisDrone with YOLOv8-S as the baseline.
- Abstract says code 'will be avaliable' with no confirmed release statement in the text.
- Reported AP on some tables is inconsistent with text (e.g., FBRT-YOLO-X AP50 48.4 is below YOLOv10-L AP50 comparisons on the same table).
Experimental setup
Datasets: VisDrone UAVDT AI-TOD
Benchmarks: VisDrone UAVDT AI-TOD
Metrics: AP AP50 AP75 Params FLOPs FPS
Baselines: YOLOv8 YOLOv9 YOLOv10 YOLOv5 RT-DETR DMNet QueryDet CEASC YOLC GLSAN GFL ClusDet
Builds on: YOLOv8 FPN PANet RT-DETR
Closest papers by meaning
- RemDet: Rethinking Efficient Model Design for UAV Object Detection (AAAI 2025 · similarity 0.87)
- Mamba YOLO: A Simple Baseline for Object Detection with State Space Model (AAAI 2025 · similarity 0.85)
- FRED: Towards a Full Rotation-Equivariance in Aerial Image Object Detection (AAAI 2024 · similarity 0.85)
- Multiview Aerial Visual RECognition (MAVREC): Can Multi-view Improve Aerial Visual Perception? (CVPR 2024 · similarity 0.85)
- Learning 1-Bit Tiny Object Detector with Discriminative Feature Refinement (ICML 2024 · similarity 0.83)
- PointOBB: Learning Oriented Object Detection via Single Point Supervision (CVPR 2024 · similarity 0.83)
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.