MODULE_TODO.md

KELPIE — Design & Implementation TODO

Paper: Prototype-Based Low Altitude UAV Semantic Segmentation (PBSeg) ArXiv: https://arxiv.org/abs/2604.01550 Defense Score: 36/50 | Tier: T2

Paper Verification

• Read paper fully
• Clone reference repo
• Check repo runnability (blocked: upstream repo empty)
• Identify gaps vs our hardware
• Document take / skip / adapt
• Verdict: proceed with CTO review flag

Architecture Design

• Define input/output interfaces
• Map dependencies to other wave modules
• Choose baseline backbone and PBCA design
• Design config schema
• Write architecture documentation

Phase 1 — Scaffold

• pyproject.toml dependencies
• src/anima_kelpie/ package layout
• Config loading (TOML + Pydantic)
• Synthetic dataset fixture
• Unit test skeleton
• Docker setup verification in runtime

Phase 2 — Paper Reproduction

• Acquire UAVid and UDD6 datasets
• Implement exact preprocessing pipeline from paper assumptions
• Train on reference datasets
• Evaluate and compare to paper claims
• Document results in NEXT_STEPS.md

Phase 3 — Hardware Adaptation

• Integrate internal 1.8M UAV dataset pipeline
• MLX backend parity improvements
• CUDA optimization on Vast.ai GPU
• Live inference benchmark (latency/FPS)

Phase 4 — Integration

• ROS2 topic node
• Gazebo bridge integration
• Dockerized service validation
• API and integration tests
• Shenzhen demo readiness checklist