ASAP Bird Detection

ASAP (Asynchronous Slicing Accelerated Pipeline) runtime code for high-resolution small-bird detection experiments.

What this repository supports

• Public demo commands for user-supplied local image/video inputs
• Core ASAP image/video inference pipeline
• TensorRT export helper for rebuilding machine-specific engines
• Lightweight environment checks and unit tests

Public/private asset boundary

• The paper's internal private 4K surveillance dataset is not published in this repository.
• The local data/ directory, including data/samples/, is not part of the public repository contract.
• Model weights, TensorRT engines, generated outputs, paper figures, and raw experiment logs are not included.
• Paper benchmark reference numbers are documented in PAPER_RESULTS.md; exact paper-number regeneration is intentionally not packaged as public tooling because it depends on private assets and matching hardware/software setup.

Quick Start

cd <repo-root>
python -m pip install -r requirements.txt
python main.py doctor

TensorRT export dependencies are split out because they are only needed on CUDA/TensorRT machines:

python -m pip install -r requirements-engine.txt

Run a demo with your own input

python main.py demo video -i /path/to/video.mp4
python main.py demo image -i /path/to/image.jpg --save -o outputs/runtime

Default sample paths are local workspace conveniences only; public clones should pass explicit -i/--input paths.

Advanced runtime commands

python main.py video -i /path/to/video.mp4 --model /path/to/yolo11n_1280.engine
python main.py image -i /path/to/images --model /path/to/yolo11n_1280.engine --save
python main.py export -m /path/to/yolo11n.pt --imgsz 1280 --batch 16

Paper-aligned TensorRT engines should be rebuilt on the target machine as dynamic FP16 engines with max batch 16. The paper's "8 patches" row refers to the 4K tiling count, not the TensorRT engine max batch.

Verification

bash scripts/verify_public_surface.sh

Reference files

• PAPER_RESULTS.md — paper benchmark reference values and public/private boundary

Local-only files

Do not commit datasets, model weights, TensorRT engines, generated outputs, paper figures, raw experiment logs, local environment files, credentials, or personal integration code. The repository .gitignore covers the usual paths and file types.

License

This project is released under the GNU Affero General Public License v3.0 (AGPL-3.0-only). It uses Ultralytics YOLO tooling, which is available under AGPL-3.0 or a commercial Ultralytics Enterprise license.

Citation

If you use this repository, cite the ASAP paper/manuscript:

• CITATION.cff
• Paper title: Asynchronous Slicing Accelerated Pipeline for Real-time High-Resolution Small Bird Flock Monitoring

Use the final publication metadata when the paper is formally published.