Skip to content

PRD.md

Latest commit

 

History

History
94 lines (78 loc) · 3.75 KB

PRD.md

File metadata and controls

94 lines (78 loc) · 3.75 KB
Raw

SLOUGHI — Implementation PRD (Wave-10 WARDOG)

1. Module Identity

2. Verification Verdict (Gate 0)

Status: VERIFIED_WITH_RISKS

Completed:

  • Paper read end-to-end (papers/2503.06625.pdf, extracted text in papers/2503.06625.txt)
  • Official repo confirmed and cloned (references/SGLATrack)
  • Core formulas and architecture extracted (paper §3.1-§3.3, Eq. 4-8)
  • Hyperparameters extracted (paper §4.1 + reference yaml)
  • Reported metrics cross-checked against upstream README tables
  • Independent signal checked (OpenAlex indexed, low citation count)

Risks (must be tracked):

  • Required datasets not present under expected shared-volume paths.
  • Upstream environment is pinned to legacy stack and is not plug-and-play in local environment.
  • Upstream issue tracker includes unresolved dataset/env/FPS discrepancies.
  • Public benchmark links are partially unstable.

If these risks block full reproduction in Phase 2/3, escalate for CTO review before claims.

3. What We Take / Skip / Adapt

Take

  • One-stream ViT tracker design.
  • Similarity-guided layer adaptation mechanism.
  • Selection module and layer-wise similarity supervision.

Skip

  • Non-UAV deployment assumptions not relevant to ANIMA stack.
  • Any dataset pipeline that cannot be legally or operationally sourced.

Adapt

  • Interface contracts for YOLO26 detector priors.
  • Dual runtime backend selection and tolerance guardrails.
  • Internal 1.8M UAV dataset ingestion once gate opens.

4. Technical Core (Paper-Aligned)

Inputs

  • Template image: 128x128
  • Search image: 256x256

Model Structure

  • Patch embedding for template/search
  • One-stream token fusion
  • Saturation-aware layer adaptation
  • Selection module outputs candidate layer probabilities
  • Retain selected post-saturation layer(s), disable redundant layers
  • Tracking head predicts box and confidence

Loss

  • Tracking objective (L1 + IoU surrogate + classification)
  • Similarity-guided selection objective (Lsim per paper Eq. 6/7 style)
  • Combined form following paper Eq. (8) coefficients

Default Hyperparameters (from paper)

  • Optimizer: AdamW
  • Epochs: 300
  • LR: 4e-4 (head), 4e-5 (backbone/selector)
  • Batch size: 32
  • LR drop: epoch 240
  • l* = 6 selected as best speed/accuracy trade-off in ablation

5. Benchmarks and Targets

Primary paper targets:

  • UAV123 AUC: 66.9 (SGLATrack-DeiT*)
  • UAV123 speed: 224.7 FPS GPU / 74.8 FPS CPU
  • DTB70, UAVDT, UAVTrack112, UAVTrack112_L AUC levels as in paper Table 1

Project targets (initial stage):

  • Scaffold parity: shape-correct forward + stable CLI + deterministic tests
  • Benchmark harness: runnable on synthetic data now, real datasets when available
  • Reproduction target: within ±5% of paper metrics once datasets/checkpoints are available

6. Work Packages

Detailed breakdown is now managed in:

  • ASSETS.md (assets + risk register)
  • prds/README.md and prds/PRD-01..07
  • tasks/INDEX.md and tasks/PRD-xxxx.md

7. Exit Criteria for Initial Stage

  • Full PRD/task suite exists and is consistent with paper
  • Initial SGLATrack scaffold code runs on CPU path
  • Dual backend flags (mlx|cuda|cpu) are wired into all scripts
  • YOLO26 integration contract is documented and validated in CLI schema
  • Verification risks are recorded in NEXT_STEPS.md