Skip to content

ASSETS.md

Latest commit

 

History

History
66 lines (52 loc) · 3.39 KB

ASSETS.md

File metadata and controls

66 lines (52 loc) · 3.39 KB
Raw

WHIPPET — Asset Manifest

Paper

  • Title: EDNet: Edge-Optimized Small Target Detection in UAV Imagery - Faster Context Attention, Better Feature Fusion, and Hardware Acceleration
  • ArXiv: 2501.05885
  • Correct arXiv date: 2025-01-10
  • Authors: Zhifan Song, Yuan Zhang, Abd Al Rahman M. Abu Ebayyeh

Status

PROCEED-WITH-CAVEATS

Pretrained Weights

Model Size Source Path / URL Status
EDNet-Tiny 4.0 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/tiny.pt READY
EDNet-N 6.3 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/nano.pt READY
EDNet-S 19.2 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/small.pt READY
EDNet-M 77.6 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/medium.pt READY
EDNet-B 51.7 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/big.pt READY
EDNet-L 64.3 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/large.pt READY
EDNet-X 98.4 MB Upstream repo https://raw.githubusercontent.com/zsniko/EDNet/main/pretrained/xlarge.pt READY
YOLO26m base unknown Ultralytics docs yolo26m.pt UNVERIFIED-LOCAL
YOLO26m UAV finetune internal ANIMA requirement yolo26m-uav.pt MISSING

Datasets

Dataset Size Split Source Path Status
VisDrone 2019-DET 6471 train / 548 val / 1610 test-dev train/val/test VisDrone /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/visdrone FOUND
UAVDT not yet confirmed train/test UAVDT /Volumes/AIFlowDev/RobotFlowLabs/datasets/... MISSING
DroneVehicle not yet confirmed paper extension only DroneVehicle /Volumes/AIFlowDev/RobotFlowLabs/datasets/... MISSING
SeaDronesSee not yet confirmed paper extension only SeaDronesSee /Volumes/AIFlowDev/RobotFlowLabs/datasets/... MISSING
1.8M Mega UAV internal train/val/test internal canonical path TBD MISSING

Hyperparameters From Paper

Param Value Paper Section
optimizer SGD III.B.1
learning_rate 0.01 III.B.1
momentum 0.9 III.B.1
epochs 200 III.B.1
image_size 640 repo README / model table
loss WIoUv3 II.C
training hardware NVIDIA A100 80GB PCIe Table I

Expected Metrics

Benchmark Metric Paper Value Our Target
VisDrone val EDNet-Tiny mAP50 33.3-34.1 depending on source table >= 33.0 in reproduction
VisDrone val EDNet-M mAP50 47.1 >= 45.5 in first reproduction
VisDrone val EDNet-B mAP50 48.3-48.5 depending on source table >= 47.0 in first reproduction
VisDrone val EDNet-X mAP50 50.2-50.6 depending on source table >= 49.0 in first reproduction
iPhone 12 runtime 16-55 FPS paper-reported only until reproduced

Notes

  • The paper and repo contain minor metric presentation drift between the PDF table and the repo README. Treat the PDF as primary and the repo as a helpful secondary source.
  • YOLO26 assets are an ANIMA requirement and must be verified in the runtime environment before training scripts depend on them.