ASSETS.md

BEAGLE — Asset Manifest

Paper

• Title: SWA-PF: Semantic-Weighted Adaptive Particle Filter for Memory-Efficient 4-DoF UAV Localization in GNSS-Denied Environments
• ArXiv: 2509.13795
• Authors: Jiayu Yuan, Ming Dai, Enhui Zheng, Chao Su, Nanxing Chen, Qiming Hu, Shibo Zhu, Yibin Cao
• Status: ALMOST

Reference Implementation

Asset	Source	Local Path	Status
Public repo	GitHub	repositories/SWA-PF	DONE
Paper PDF	arXiv/local	papers/2509.13795.pdf	DONE
Independent reproduction	public web	—	MISSING

Pretrained Weights

Model	Purpose	Source	Expected Path	Status
SegFormer-B0	UAV semantic segmentation	paper/repo logs	artifacts/segformer_b0_uav.pth	MISSING
U-Net VGG	satellite semantic segmentation	paper/repo training output	artifacts/unet_vgg_satellite.pth	MISSING
YOLO26m	adaptation-only target prior	internal Ultralytics stack	artifacts/yolo26m-uav.pt	MISSING

Datasets

Dataset	Purpose	Source	Expected Path	Status
MAFS	paper route benchmark	Baidu share / internal mirror	/Volumes/AIFlowDev/RobotFlowLabs/datasets/MAFS	MISSING
SemanticMAFS	paper semantic supervision	derived from MAFS	/Volumes/AIFlowDev/RobotFlowLabs/datasets/SemanticMAFS	MISSING
VisDrone	later UAV adaptation	shared volume	/Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/visdrone	DONE
UAVDT	later UAV adaptation	public	/Volumes/AIFlowDev/RobotFlowLabs/datasets/UAVDT	MISSING
DroneVehicle	later multimodal adaptation	public	/Volumes/AIFlowDev/RobotFlowLabs/datasets/DroneVehicle	MISSING
SeaDronesSee	later maritime adaptation	public	/Volumes/AIFlowDev/RobotFlowLabs/datasets/SeaDronesSee	MISSING
1.8M Mega UAV	internal defense training set	internal	/Volumes/AIFlowDev/RobotFlowLabs/datasets/mega_uav_1p8m	UNKNOWN

Hyperparameters From Paper

Param	Value	Paper Reference
UAV semantic model	SegFormer-B0	§IV.A.2
Satellite semantic model	VGG-pretrained U-Net	§IV.A.1
UAV semantic input size	512x512	§IV.A.2
UAV freeze epochs	50	§IV.A.2
UAV finetune epochs	200	§IV.A.2
UAV freeze batch size	32	§IV.A.2
UAV finetune batch size	8	§IV.A.2
Optimizer	AdamW	§IV.A.2
Learning rate	1e-4	§IV.A.2
Weight decay	1e-2	§IV.A.2
Scheduler	cosine annealing	§IV.A.2
Particle resize target	400x400	§IV.B.3
Rotation bins	100	§IV.C
Motion noise epsilon	15	§V.A.1
Gamma	10	§V.A.1
Fixed-altitude particles	5000	§V.A.1
Variable-altitude particles	40000	§V.A.1

Expected Metrics

Benchmark	Metric	Paper Value	Our Target
MAFS-10	RMSE	6.5685 m	<= 7.0 m
MAFS-10	Recall@10	97.368%	>= 95%
MAFS-10	Median error	6.653 m	<= 7.0 m
MAFS-10	Fitting time	7 s	<= 10 s
MAFS-10	Finish time	25 s	<= 35 s

Notes

• Exact paper reproduction is blocked on MAFS ingestion and missing semantic model weights.
• The public repo is a valid algorithmic reference, not a production baseline.
• YOLO26 integration is a future adaptation task and is intentionally not treated as part of the paper reproduction scope.