55 lines (48 loc) · 3.28 KB

TERRIER — Asset Manifest

Paper

Title: Progressive Representation Learning for Real-Time UAV Tracking
ArXiv: https://arxiv.org/abs/2409.16652
DOI: https://doi.org/10.1109/IROS58592.2024.10803050
Authors: Changhong Fu, Xiang Lei, Haobo Zuo, Liangliang Yao, Guangze Zheng, Jia Pan

Status

Module state: ALMOST (verified and scaffolded; full dataset/training still pending)
Verification date: 2026-04-10

Pretrained Weights

Model	Size	Source	Path on Server	Status
AlexNet backbone (`alexnet-bn.pth`)	~233 MB	upstream PRL/HiFT dependencies	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/models/alexnet-bn.pth`	MISSING
PRL-Track snapshot	unknown	https://drive.google.com/drive/folders/1WYQf_zAMy9Xf1tLH1MRmQELe5ywwsB5d	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/models/prl-track/`	MISSING
YOLO26m base	TBD	internal YOLO26 release channel	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/models/yolo26/yolo26m.pt`	MISSING

Datasets

Dataset	Size	Split	Source	Path	Status
COCO	large	train	https://cocodataset.org/	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/coco`	DONE
GOT-10K	large	train	http://got-10k.aitestunion.com/downloads	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/got10k`	MISSING
LaSOT	large	train	http://vision.cs.stonybrook.edu/~lasot/	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/lasot`	MISSING
VisDrone	large	train/val/test	https://github.com/VisDrone/VisDrone-Dataset	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/visdrone`	PARTIAL
UAVDT	medium	benchmark	https://sites.google.com/view/daweidu/projects/uavdt	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/uavdt`	MISSING
DroneVehicle	medium	benchmark	https://github.com/VisDrone/DroneVehicle	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/dronevehicle`	MISSING
SeaDronesSee	medium	benchmark	https://seadronessee.cs.uni-tuebingen.de/	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/seadronessee`	MISSING
1.8M Mega UAV (internal)	~1.8M imgs	train/val	internal	`/Volumes/AIFlowDev/RobotFlowLabs/datasets/shared/mega_uav_1_8m`	MISSING

Hyperparameters (Paper + Upstream Config)

Param	Value	Source
template size	127x127	paper Sec. IV-A
search size	287x287	paper Sec. IV-A
optimizer	SGD momentum 0.9, wd 1e-4	upstream `pysot/core/config.py`
LR schedule	warmup 5e-3→1e-2 then log decay to 1e-4	paper + `experiments/config.yaml`
epochs	70 (paper text) / 100 (upstream config)	discrepancy noted
batch size	100–128 (2 GPUs)	upstream config variants

Expected Metrics (from paper)

Benchmark	Metric	Paper Value	Our Target
UAVTrack112	Precision	0.786	>=0.75
UAVTrack112	Success	0.602	>=0.57
UAVTrack112_L	Precision	0.803	>=0.76
UAVTrack112_L	Success	0.597	>=0.56
UAV123	Precision	0.593	>=0.56
UAV123	Success (AUC)	0.791	>=0.75
Edge platform	FPS	42.6	>=35 (target hardware)

Notes

Upstream code path is CUDA-bound in model initialization; direct CPU/MLX path needs adaptation.
No destructive download behavior should be introduced in bootstrap scripts.