ASSETS.md

FOXHOUND — Asset Manifest

Paper

• Title: Learning Motion Blur Robust Vision Transformers for Real-Time UAV Tracking
• ArXiv: 2407.05383
• URL: https://arxiv.org/abs/2407.05383
• Authors: You Wu, Xucheng Wang, Dan Zeng, Hengzhou Ye, Xiaolan Xie, Qijun Zhao, Shuiwang Li

Status

ALMOST

Why not READY:

• The public repo is present but not reproducible.
• Shared dataset volume audit did not confirm the paper datasets locally.
• Exact training recipe inherits from Aba-ViTrack and must still be recovered.

Pretrained Weights

Model	Size	Source	Path on Server	Status
deit_tiny_patch16_224	~5.8M params	timm / DeiT upstream	/mnt/forge-data/models/deit_tiny_patch16_224.safetensors	MISSING
vit_tiny_patch16_224	~5.7M params	timm / ViT upstream	/mnt/forge-data/models/vit_tiny_patch16_224.safetensors	MISSING
t2t_vit_t_14	lightweight ViT	T2T-ViT upstream	/mnt/forge-data/models/t2t_vit_t_14.safetensors	MISSING
yolo26m.pt	detector backbone	internal YOLO26 baseline	/mnt/forge-data/models/yolo26/yolo26m.pt	MISSING
yolo26m-uav.pt	detector adaptation	internal YOLO26 UAV fine-tune	/mnt/forge-data/models/yolo26/yolo26m-uav.pt	MISSING

Datasets

Dataset	Size	Split	Source	Path	Status
MegaUAV-1.8M	~1.8M images	train/val/test	internal	/Volumes/AIFlowDev/RobotFlowLabs/datasets/MegaUAV-1.8M	MISSING
UAV123	benchmark	test	public benchmark	/Volumes/AIFlowDev/RobotFlowLabs/datasets/UAV123	MISSING
UAV123@10fps	derived benchmark	test	downsampled from UAV123	/Volumes/AIFlowDev/RobotFlowLabs/datasets/UAV123_10fps	MISSING
VisDrone2018	benchmark	test	public benchmark	/Volumes/AIFlowDev/RobotFlowLabs/datasets/VisDrone2018	MISSING
UAVDT	benchmark	test	public benchmark	/Volumes/AIFlowDev/RobotFlowLabs/datasets/UAVDT	MISSING
DroneVehicle	adaptation	train/val	public benchmark	/Volumes/AIFlowDev/RobotFlowLabs/datasets/DroneVehicle	MISSING
SeaDronesSee	adaptation	train/val	public benchmark	/Volumes/AIFlowDev/RobotFlowLabs/datasets/SeaDronesSee	MISSING
UAVTrack112_L	real-world eval	eval	paper real-world set	/Volumes/AIFlowDev/RobotFlowLabs/datasets/UAVTrack112_L	MISSING

Hyperparameters From Paper

Param	Value	Paper Section
template_size	128 x 128	§4.1
search_size	256 x 256	§4.1
loss_cls	weighted focal loss	§3.4
loss_box	L1 + GIoU	§3.4
eta_iou	2.0	Eq. 7
eta_l1	5.0	Eq. 7
rho	1e-4	Eq. 7, Table 5
gamma	1e3	Eq. 7
epsilon	0.05	§3.3
exit_threshold	0.95	§3.3

Hyperparameters Still Missing

Param	Status	Note
optimizer	MISSING	paper says training pipeline follows Aba-ViTrack
batch_size	MISSING	not restated in BDTrack text
schedule	MISSING	not restated in BDTrack text
nenf	MISSING	concept described, exact deployed value not exposed in paper text
lambda_weight	MISSING	equation includes lambda but text does not provide the final scalar
tau	MISSING	equation includes tau but final chosen constant is not stated in exposed paper text

Expected Metrics

Benchmark	Metric	Paper Value	Our Target
four-benchmark average	precision	84.4	>= 83.0 bootstrap reproduction
four-benchmark average	success	64.5	>= 63.0 bootstrap reproduction
UAVDT	precision	84.1	>= 82.0
UAVDT	success	61.0	>= 59.0
VisDrone2018	precision	85.2	>= 83.0
VisDrone2018	success	64.3	>= 62.0
UAV123	precision	84.8	>= 83.0
UAV123	success	66.7	>= 65.0
UAV123@10fps	precision	83.5	>= 82.0
UAV123@10fps	success	65.9	>= 64.0
efficiency	GPU FPS	283.4	>= 240 on modern CUDA path

Reproducibility Notes

• Public GitHub repo currently lacks the implementation needed for a direct reproduction.
• Paper implementation details partially inherit from Aba-ViTrack, so FOXHOUND must recover or re-derive missing training settings before Phase 2.
• YOLO26 is an adaptation layer for ANIMA deployment, not part of the original BDTrack paper.