HazyDet
收藏资源简介:
HazyDet是由解放军工程大学等机构创建的一个大规模数据集,专门用于雾霾场景下的无人机视角物体检测。该数据集包含383,000个真实世界实例,收集自自然雾霾环境和正常场景中人工添加的雾霾效果,以模拟恶劣天气条件。数据集的创建过程结合了深度估计和大气散射模型,确保了数据的真实性和多样性。HazyDet主要应用于无人机在恶劣天气条件下的物体检测,旨在提高无人机在复杂环境中的感知能力。
HazyDet is a large-scale dataset developed by institutions including the PLA Engineering University, specially tailored for object detection from unmanned aerial vehicle (UAV) perspectives in hazy scenarios. The dataset comprises 383,000 real-world instances collected from both natural hazy environments and normal scenes with artificially added haze effects to simulate adverse weather conditions. Its development process integrates depth estimation and atmospheric scattering models to ensure the authenticity and diversity of the dataset. HazyDet is primarily applied to UAV-based object detection under adverse weather conditions, aiming to enhance the perception capabilities of UAVs in complex environments.
HazyDet: Open-source Benchmark for Drone-View Object Detection with Depth-cues in Hazy Scenes
数据集概述
HazyDet-365K
-
下载地址: Baidu Cloud
-
数据结构:
HazyDet-365K |-- train |-- clean images |-- hazy images |-- labels |-- val |-- clean images |-- hazy images |-- labels |-- test |-- clean images |-- hazy images |-- labels |-- RDDTS |-- hazy images |-- labels
-
密码:
grok
模型与性能
检测器 (Detectors)
| 模型 | 骨干网络 | 参数数量 (M) | GFLOPs | mAP on Test-set | mAP on RDDTS | 配置文件 | 权重文件 |
|---|---|---|---|---|---|---|---|
| YOLOv3 | Darknet53 | 61.63 | 20.19 | 35.0 | 19.2 | config | weight |
| GFL | ResNet50 | 32.26 | 198.65 | 36.8 | 13.9 | config | weight |
| YOLOX | CSPDarkNet | 8.94 | 13.32 | 42.3 | 24.7 | config | weight |
| RepPoints | ResNet50 | 36.83 | 184.32 | 43.8 | 21.3 | config | weight |
| FCOS | ResNet50 | 32.11 | 191.48 | 45.9 | 22.8 | config | weight |
| Centernet | ResNet50 | 32.11 | 191.49 | 47.2 | 23.8 | config | weight |
| ATTS | ResNet50 | 32.12 | 195.58 | 50.4 | 25.1 | config | weight |
| DDOD | ResNet50 | 32.20 | 173.05 | 50.7 | 26.1 | config | weight |
| VFNet | ResNet50 | 32.89 | 187.39 | 51.1 | 25.6 | config | weight |
| TOOD | ResNet50 | 32.02 | 192.51 | 51.4 | 25.8 | config | weight |
| Sparse RCNN | ResNet50 | 108.54 | 147.45 | 27.7 | 10.4 | config | weight |
| Dynamic RCNN | ResNet50 | 41.35 | 201.72 | 47.6 | 22.5 | config | weight |
| Faster RCNN | ResNet50 | 41.35 | 201.72 | 48.7 | 23.6 | config | weight |
| Libra RCNN | ResNet50 | 41.62 | 209.92 | 49.0 | 23.7 | config | weight |
| Grid RCNN | ResNet50 | 64.46 | 317.44 | 50.5 | 25.2 | config | weight |
| Cascade RCNN | ResNet50 | 69.15 | 230.40 | 51.6 | 26.0 | config | weight |
| Conditional DETR | ResNet50 | 43.55 | 94.17 | 30.5 | 11.7 | config | weight |
| DAB DETR | ResNet50 | 43.70 | 97.02 | 31.3 | 11.7 | config | weight |
| Deform DETR | ResNet50 | 40.01 | 192.51 | 51.9 | 26.5 | config | weight |
| FCOS-DeCoDet | ResNet50 | 34.62 | 225.37 | 47.4 | 24.3 | config | weight |
| VFNet-DeCoDet | ResNet50 | 34.61 | 249.91 | 51.5 | 25.9 | config | weight |
去雾 (Dehazing)
| 类型 | 方法 | PSNR | SSIM | mAP on Test-set | mAP on RDDTS | 权重文件 |
|---|---|---|---|---|---|---|
| Baseline | Faster RCNN | - | - | 39.5 | 21.5 | weight |
| Dehaze | GridDehaze | 12.66 | 0.713 | 38.9 (-0.6) | 19.6 (-1.9) | weight |
| Dehaze | MixDehazeNet | 15.52 | 0.743 | 39.9 (+0.4) | 21.2 (-0.3) | weight |
| Dehaze | DSANet | 19.01 | 0.751 | 40.8 (+1.3) | 22.4 (+0.9) | weight |
| Dehaze | FFA | 19.25 | 0.798 | 41.2 (+1.7) | 22.0 (+0.5) | weight |
| Dehaze | DehazeFormer | 17.53 | 0.802 | 42.5 (+3.0) | 21.9 (+0.4) | weight |
| Dehaze | gUNet | 19.49 | 0.822 | 42.7 (+3.2) | 22.2 (+0.7) | weight |
| Dehaze | C2PNet | 21.31 | 0.832 | 42.9 (+3.4) | 22.4 (+0.9) | weight |
| Dehaze | DCP | 16.98 | 0.824 | 44.0 (+4.5) | 20.6 (-0.9) | weight |
| Dehaze | RIDCP | 16.15 | 0.718 | 44.8 (+5.3) | 24.2 (+2.7) | weight |

- 1HazyDet: Open-source Benchmark for Drone-view Object Detection with Depth-cues in Hazy Scenes石家庄校区,解放军工程大学 · 2024年



