Research map / Detection and segmentation / CNN features
RemDet: Rethinking Efficient Model Design for UAV Object 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 RemDet, a real-time one-stage anchor-free UAV object detector built from ChannelC2f, a multiplication-based GatedFFN, and a Context Enhanced Downsample (CED) module, guided by an information-loss design principle.
Problem
Real-time object detection in UAV images, where objects are small and dense and compute is constrained; existing real-time detectors are not optimized for UAV imagery and small-object methods are heavily handcrafted and slow.
Approach
Use information loss as a design guideline: widen intra-layer representations (ChannelC2f), replace MLP/FFN with cheap element-wise multiplication (GatedFFN) for high-dimensional features, and downsample with depthwise convolution plus Patch Merge (CED) to reduce information loss, avoiding heavy handcrafted modules.
Key results
- RemDet-X with cluster-aware cropping reaches 40.0 AP on VisDrone validation, versus 31.8 for YOLC, with 9.0 ms latency on a single RTX 4090.
- RemDet-Tiny reaches 21.8 AP on VisDrone at 3.4 ms latency and 4.6 GFLOPs, above YOLOv8-N (19.1 AP).
- On UAVDT, RemDet-L (with cluster-aware cropping) reaches 20.6 AP versus 19.3 for YOLC, the best prior result in that table.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GatedFFN (ablation) | YOLOv8-N on VisDrone · VisDrone val | AP95: 19.1 -> 20.3; ↑ 1.2 | Helped |
| add CED (ablation) | YOLOv8-N on VisDrone · VisDrone val | AP95: 19.1 -> 19.8; ↑ 0.7 | Helped |
| add ChannelC2f (ablation) | YOLOv8-N on VisDrone · VisDrone val | AP95: 19.1 -> 19.9; ↑ 0.8 | Helped |
| add GatedFFN and CED (ablation) | YOLOv8-N on VisDrone · VisDrone val | AP95: 19.1 -> 21.5; ↑ 2.4 | Helped |
| add ChannelC2f to GatedFFN and CED (ablation) | YOLOv8-N on VisDrone · VisDrone val | AP95: 21.5 -> 21.7; ↑ 2.6 | Helped |
| increase CED channel expansion ratio from 1 to 2 (ablation) | RemDet · UAV detection | AP | Helped |
| remove channel expansion from stages other than stage 1 (ablation) | RemDet · UAV detection | AP: 0.1% | Hurt |
| use block ratio (3:6:6:3) rather than (3:3:6:3) (ablation) | RemDet · VisDrone | accuracy: 20.3% | Hurt |
| use multiplication instead of MLP at comparable computational cost (ablation) | RemDet GatedFFN · UAV detection | mAP: 0.3% | Helped |
| retain the w2^Tx term in multiplication (ablation) | Multiplication module · UAV detection | accuracy: 0.4% | Helped |
| replace CSPDarknet with RemDet backbone | YOLOv5-S on VisDrone · VisDrone val | AP95: 18.6 -> 20.7; ↑ 2.1 | Helped |
| replace CSPNeXt with RemDet backbone | RTMDet on VisDrone · VisDrone val | AP95: 21.8 -> 22.8; ↑ 1.0 | Helped |
Surprising findings
- Multiplication-based GatedFFN matches or exceeds MLP-style feed-forward accuracy at comparable cost, and the paper argues the multiplicative form is more cost-effective for high-dimensional representation.
- RemDet is reported to be weaker than competing methods on large objects, which the authors attribute to heavily handcrafted designs in those methods.
Limitations
- Accuracy on large objects lags behind handcrafted methods on VisDrone and UAVDT.
- The paper states the model is superior only in the tiny version on MSCOCO, with accuracy gaps at larger scales.
- Some comparisons use different hardware, input sizes, and loss/FLOPs counting tools (e.g., thop vs fvcore), so fairness is not fully established.
Experimental setup
Datasets: VisDrone UAVDT MSCOCO
Benchmarks: VisDrone UAVDT MSCOCO
Metrics: AP AP50 AP75 APs APm APl latency (ms) FLOPs (G) FPS
Baselines: YOLOv8 YOLOv6-v3.0 YOLOv7-Tiny RTMDet QueryDet CEASC YOLC CZDet UFPMP-Det GOLD-YOLO YOLOv10
Builds on: YOLOv8 C2f ConvNeXt Gated Linear Units Cluster-Aware Crops
Closest papers by meaning
- FBRT-YOLO: Faster and Better for Real-Time Aerial Image Detection (AAAI 2025 · similarity 0.87)
- Multiview Aerial Visual RECognition (MAVREC): Can Multi-view Improve Aerial Visual Perception? (CVPR 2024 · similarity 0.86)
- PointSR: Self-Regularized Point Supervision for Drone-View Object Detection (CVPR 2025 · similarity 0.84)
- Detection-Friendly Nonuniformity Correction: A Union Framework for Infrared UAV Target Detection (CVPR 2025 · similarity 0.84)
- UAV3D: A Large-scale 3D Perception Benchmark for Unmanned Aerial Vehicles (NeurIPS 2024 · similarity 0.84)
- Learning 1-Bit Tiny Object Detector with Discriminative Feature Refinement (ICML 2024 · similarity 0.84)
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.