Skip to content

PRD.md

Latest commit

 

History

History
49 lines (41 loc) · 2.63 KB

PRD.md

File metadata and controls

49 lines (41 loc) · 2.63 KB

COONHOUND — MambaNUT Build Manual

Verification Verdict

  • Paper: real, arXiv 2412.00626v3, dated 2025-05-10
  • Venue claim: matches IROS 2025
  • Upstream repo: exists and matches the paper naming/configuration
  • Dataset status: partial; visdrone and uavdark135 are on the shared volume, uavdt is not present yet
  • Reproducibility status: VERIFIED WITH CAVEATS

Why The Verdict Is Not "Clean"

  • The upstream repo contains a hardcoded pretrained Mamba path.
  • Nighttime datasets and released checkpoints are distributed via Baidu links, which is fragile for reproducibility.
  • The README has setup inconsistencies, including requirement.txt vs requirements.txt.
  • No independent reproduction report was confirmed during the initial verification pass.

What COONHOUND Builds

  • Core method retained: Vision Mamba backbone plus center head with Adaptive Curriculum Learning.
  • Adaptation path added: YOLO26 detector bootstrap before Mamba-based nighttime tracking.
  • Deployment target retained: dual compute execution on MLX, CUDA, and CPU fallback.

What We Take

  • One-stream tracker structure.
  • mambar_small_patch16_224 backbone profile.
  • ACL sampling scheduler and ADW loss formulation.
  • Paper training recipe and benchmark targets.

What We Skip

  • Direct dependence on the upstream absolute file layout.
  • Baidu-only download automation.
  • Full training or evaluation claims before the missing datasets and weights are staged.

What We Adapt

  • Detector-first pipeline with YOLO26 ROI bootstrap.
  • Internal 1.8M UAV dataset and shared-wave datasets.
  • Clean local configuration, asset checks, and task decomposition for ANIMA.

Outputs Created In This Pass

Immediate Build Order

  1. Land clean foundation/config/runtime checks.
  2. Port the ACL math and dataset mixture logic exactly.
  3. Implement the Vision Mamba tracker module behind a backend-agnostic adapter.
  4. Attach YOLO26 proposals to template/search crop generation.
  5. Stage missing datasets and pretrained weights.
  6. Run reproduction on reference benchmarks before Shenzhen-specific fine-tuning.