ANIMA WHIPPET — EDNet on YOLO26

WHIPPET is the Wave-10 WARDOG module for paper-grounded small-target UAV detection. The upstream paper is EDNet, a 2025 YOLOv10-based detector for VisDrone. This module does not treat YOLO26 as part of the paper. It treats YOLO26 as the required ANIMA rebase target for production.

Current Status

• Paper: verified as real, dated 2025-01-10 on arXiv.
• Upstream repo: public and populated with code, configs, and pretrained weights.
• Dataset coverage: VisDrone present on the shared volume; UAVDT and internal mega-UAV paths still need confirmation.
• Build state: initial PRD suite, task breakdown, paper blocks, and YOLO26 rebase scaffold are implemented.

Key Verification Findings

• The local metadata originally listed the paper as 2026-04-10; the actual arXiv paper is 2025-01-10.
• The paper is built on YOLOv10 and reports VisDrone results. YOLO26 is an ANIMA deployment requirement, not part of the original method.
• The paper reports training on VisDrone 2019-DET with SGD, lr=0.01, momentum=0.9, and 200 epochs on an A100 80GB system.
• Upstream reproducibility looks plausible because the repo publishes model configs, pretrained weights, and a dataset YAML. Independent third-party reproduction is still limited, so this module proceeds with caution.

Repo Layout

project_whippet/
├── ASSETS.md
├── PRD.md
├── NEXT_STEPS.md
├── MODULE_TODO.md
├── prds/
├── tasks/
├── configs/
├── scripts/
├── src/anima_whippet/
└── tests/

Quick Start

uv pip install -e ".[dev]"
python -m anima_whippet backend
python -m anima_whippet verify-paper
python -m anima_whippet check-assets
python -m anima_whippet export-yolo26 --scale m --output configs/models/whippet-yolo26m.yaml
pytest -q

Initial Deliverables

• ASSETS.md: paper assets, weights, datasets, and readiness gaps.
• prds/: full 7-PRD ANIMA build suite.
• tasks/: granular implementation backlog derived from the PRDs.
• src/anima_whippet/paper_blocks.py: paper-inspired FCA and C2f-FCA building blocks.
• src/anima_whippet/yolo26_rebase.py: generated YOLO26 rebase YAML scaffold.

Non-Goals for This Initial Stage

• Claiming paper reproduction before a real training run.
• Claiming YOLO26 parity with upstream EDNet before the custom modules are registered into the actual Ultralytics training stack.
• Downloading gated datasets without explicit permission.