Skip to content

ASSETS.md

Latest commit

 

History

History
84 lines (76 loc) · 4.31 KB

ASSETS.md

File metadata and controls

84 lines (76 loc) · 4.31 KB

COONHOUND — Asset Manifest

Paper

  • Title: MambaNUT: Nighttime UAV Tracking via Mamba-based Adaptive Curriculum Learning
  • ArXiv: 2412.00626
  • ArXiv version: v3
  • ArXiv version date: 2025-05-10
  • Venue: IROS 2025
  • Authors: You Wu, Xiangyang Yang, Xucheng Wang, Hengzhou Ye, Dan Zeng, Shuiwang Li

Status

  • Verification verdict: VERIFIED WITH CAVEATS
  • Independent reproductions found: None confirmed during initial scan
  • Reproduction mode for this module: integration-prep + code-grounded scaffold

Pretrained Weights

Model Size Source Expected Path Status
mambar_small_patch16_224.pth unknown Hugging Face Wangf3014/Mamba-Reg pretrained_models/mambar_small_patch16_224.pth MISSING
MambaNUT_ep0300.pth.tar unknown Baidu NetDisk (repo README) output/checkpoints/train/mambanut/mambar_small_patch16_224/ MISSING
yolo26m.pt unknown Internal YOLO26 baseline weights/yolo26m.pt MISSING
yolo26m-uav.pt unknown Internal fine-tuned UAV detector weights/yolo26m-uav.pt MISSING

Datasets Used By Paper

Dataset Role Source Local Path Candidate Status
LaSOT daytime train official external only UNCHECKED
GOT-10k daytime train/val official external only UNCHECKED
COCO 2017 daytime train official external only UNCHECKED
TrackingNet daytime train official external only UNCHECKED
BDD100K-Night nighttime train Baidu link in repo README external only MISSING
SHIFT-Night nighttime train Baidu link in repo README external only MISSING
ExDark nighttime train Baidu link in repo README external only MISSING
NAT2021 evaluation benchmark external only MISSING
NAT2024-1 evaluation benchmark external only MISSING
UAVDark135 evaluation benchmark /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/uavdark135 FOUND

Datasets Used By COONHOUND Adaptation

Dataset Role Source Local Path Candidate Status
1.8M Mega UAV primary adaptation train internal /Volumes/AIFlowDev/RobotFlowLabs/datasets/ UNCHECKED
VisDrone detection adaptation shared volume /Volumes/AIFlowDev/RobotFlowLabs/datasets/wave10_staging/visdrone FOUND
UAVDT tracking adaptation shared volume / official /Volumes/AIFlowDev/RobotFlowLabs/datasets/ MISSING
DroneVehicle cross-modal adaptation official /Volumes/AIFlowDev/RobotFlowLabs/datasets/ UNCHECKED
SeaDronesSee maritime adaptation official /Volumes/AIFlowDev/RobotFlowLabs/datasets/ UNCHECKED

Hyperparameters From Paper

Param Value Paper Evidence
optimizer AdamW Sec. IV-A
learning_rate 4e-4 Sec. IV-A
weight_decay 1e-4 Sec. IV-A
epochs 300 Sec. IV-A
lr_drop_epoch 240 Sec. IV-A
batch_size 32 Sec. IV-A
samples_per_epoch 60000 Sec. IV-A
search_size 256 Sec. IV-A
template_size 128 Sec. IV-A
backbone mambar_small_patch16_224 repo config + Sec. IV-A
backbone_embed_dim 384 repo model build
backbone_depth 24 repo model build
head_type CENTER repo config
ACL ramp constant 150 Eq. 3 text (theta = 150)
loss weights GIoU=2.0, L1=5.0, ADW=1e-5 Eq. 8 + repo actor

Expected Metrics

Benchmark Metric Paper Value COONHOUND Gate
NAT2024-1 Precision 83.3 >= 79.0
NAT2024-1 Norm. Precision 76.9 >= 73.0
NAT2024-1 Success 63.6 >= 60.0
NAT2021 Precision 70.1 >= 66.0
NAT2021 Norm. Precision 64.6 >= 61.0
NAT2021 Success 52.4 >= 49.0
UAVDark135 Precision 70.0 >= 66.0
UAVDark135 Norm. Precision 69.3 >= 65.0
UAVDark135 Success 57.1 >= 54.0
Real-world Jetson Orin NX throughput > 30 FPS >= 25 FPS

Reproducibility Caveats

  • Upstream repo hardcodes the pretrained Mamba path to /home/wy/MambaNUT-master/pretrained_models/mambar_small_patch16_224.pth.
  • The README says requirement.txt, but the repository ships requirements.txt.
  • Nighttime training assets and tracker checkpoints are linked through Baidu NetDisk only.
  • No independent third-party reproduction was confirmed during the initial scan.