HazyDet: Open-Source Benchmark for Drone-View Object Detection With Depth-Cues in Hazy Scenes (paper)

HazyDet is the first benchmark for object detection in hazy drone imagery. It couples physics-driven synthetic data with real foggy drone photos, providing a controlled yet realistic test-bed for designing haze-robust detectors.

---

Abstract

Object detection from aerial platforms under adverse atmospheric conditions, particularly haze, is paramount for robust drone autonomy. Yet, this domain remains largely underexplored, primarily hindered by the absence of specialized benchmarks. To bridge this gap, we present HazyDet, the first, large-scale benchmark specifically designed for drone-view object detection in hazy conditions. Comprising 383,000 real-world instances derived from both naturally hazy captures and synthetically hazed scenes augmented from clear images, HazyDet provides a challenging and realistic testbed for advancing detection algorithms. To address the severe visual degradation induced by haze, we propose the Depth-Conditioned Detector (DeCoDet), a novel architecture that integrates a Depth-Conditioned Kernel to dynamically modulate feature representations based on depth cues. The practical efficacy and robustness of DeCoDet are further enhanced by its training with a Progressive Domain Fine-Tuning (PDFT) strategy to navigate synthetic-to-real domain shifts, and a Scale-Invariant Refurbishment Loss (SIRLoss) to ensure resilient learning from potentially noisy depth annotations. Comprehensive empirical validation on HazyDet substantiates the superiority of our unified DeCoDet framework, which achieves state-of-the-art performance, surpassing the closest competitor by a notable +1.5% mAP on challenging real-world hazy test scenarios. Our dataset and toolkit are available at github.

HazyDet

---

📦 Dataset at a Glance

Target size buckets: Small < 0.1 % of image area , Medium 0.1–1 % , Large > 1 %

Split	#Images	#Instances	Class	Small	Medium	Large
Train	8 000	264 511	Car	159 491	77 527	5 177
			Truck	4 197	6 262	1 167
			Bus	1 990	7 879	861
Val	1 000	34 560	Car	21 051	9 881	630
			Truck	552	853	103
			Bus	243	1 122	125
Test	2 000	65 322	Car	38 910	19 860	1 256
			Truck	881	1 409	263
			Bus	473	1 991	279
Real-world Train	400	13 753	Car	5 816	6 487	695
			Truck	86	204	57
			Bus	52	256	100
Real-world Test	200	5 543	Car	2 351	2 506	365
			Truck	26	86	30
			Bus	17	107	55

---

You can also download our HazyDet dataset from Baidu Netdisk or OneDrive.

For both training and inference, the following dataset structure is required:

HazyDet/
 ├── train/
 │   └── clean images/
 │   └── hazy images/
 │   └── lables/
 ├──val/
 │   └── clean images/
 │   └── hazy images/
 │   └── lables/
 ├── test/
 │   └── clean images/
 │   └── hazy images/
 │   └── lables/
 ├── Real-world/
 │   └── train/
 │   └── test/
 │   └── lables/
 └── README.md  <-- you are here 

Note: Both passwords for BaiduYun and OneDrive is grok.

Leadboard and Model Zoo

All the weight files in the model zoo can be accessed on Baidu Cloud and OneDrive.

Detectors

Model	Backbone	#Params (M)	GFLOPs	mAP on Synthetic Test-set	mAP on Real-world Test-set	Weight
One Stage
YOLOv3	Darknet53	61.63	20.19	35.0	30.7	weight
GFL	ResNet50	32.26	198.65	36.8	32.5	weight
YOLOX	CSPDarkNet	8.94	13.32	42.3	35.4	weight
FCOS	ResNet50	32.11	191.48	45.9	32.7	weight
VFNet	ResNet50	32.71	184.32	49.5	35.6	weight
ATTS	ResNet50	32.12	195.58	50.4	36.4	weight
DDOD	ResNet50	32.20	173.05	50.7	37.1	weight
TOOD	ResNet50	32.02	192.51	51.4	36.7	weight
Two Stage
Faster RCNN	ResNet50	41.35	201.72	48.7	33.4	weight
Libra RCNN	ResNet50	41.62	209.92	49.0	34.5	weight
Grid RCNN	ResNet50	64.46	317.44	50.5	35.2	weight
Cascade RCNN	ResNet50	69.15	230.40	51.6	37.2	weight
End-to-End
Conditional DETR	ResNet50	43.55	91.47	30.5	25.8	weight
DAB DETR	ResNet50	43.7	91.02	31.3	27.2	weight
Deform DETR	ResNet50	40.01	203.11	51.5	36.9	weight
DeCoDet
DeCoDet (Ours)	ResNet50	34.62	225.37	52.0	38.7	weight

Dehazing

Type	Method	PSNR	SSIM	mAP on Test-set		mAP on RDDTS	Weight
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

Citation

If you use this toolbox or benchmark in your research, please cite this project.

@article{feng2025HazyDet,
      title={HazyDet: Open-Source Benchmark for Drone-View Object Detection with Depth-Cues in Hazy Scenes}, 
      author={Changfeng Feng and Zhenyuan Chen and Xiang Li and Chunping Wang and Jian Yang and Ming-Ming Cheng and Yimian Dai and Qiang Fu},
      year={2025},
      journal={arXiv preprint arXiv:2409.19833}, 
}

@article{zhu2021detection,
  title={Detection and tracking meet drones challenge},
  author={Zhu, Pengfei and Wen, Longyin and Du, Dawei and Bian, Xiao and Fan, Heng and Hu, Qinghua and Ling, Haibin},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  volume={44},
  number={11},
  pages={7380--7399},
  year={2021},
  publisher={IEEE}
}