SmartMedia YOLO26 Small-Object Experiments

Public, reproducible YOLO26-family small-object detection ablation code used for SmartMedia experiments. This repository keeps only the fair public surface: official YOLO26/YOLO26-P2 configs, single-axis derived variants, dataset-preparation utilities, training/validation scripts, and public-safe notes.

Fairness boundary

The public repo intentionally excludes legacy exploratory baselines that were not based on the official YOLO26 family. The included variants are derived from official Ultralytics YOLO26 config files in a controlled way:

• baseline: official yolo26.yaml.
• p2: official yolo26-p2.yaml.
• spd: official yolo26.yaml with only the first two downsampling layers changed from Conv to Focus as a space-to-depth-style stem.
• spd_p2: official yolo26-p2.yaml with the same SPD-style stem change.
• p2_nop5: official yolo26-p2.yaml with the P5 detection head removed.
• spd_p2_nop5: combines the SPD-style stem and P5-head removal changes.

This makes the comparison fair as a controlled architecture ablation. Exact reproduction of the released official yolo26n.pt checkpoint is not claimed, because the checkpoint metadata points to an intermediate pretrained model and internal training knobs. See RESULTS.md for the reproduction boundary.

Repository layout

configs/   Official YOLO26 configs and controlled derived variants
scripts/   data conversion, training, validation, ablation, and analysis utilities
results/   public-safe sanity metrics and dataset statistics

Raw datasets, checkpoints, generated runs, logs, paper drafts, slides, and private environment files are intentionally excluded.

Setup

python -m venv .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

Use a PyTorch/CUDA build that matches your machine if the default resolver is not suitable.

Prepare VisDrone

python scripts/prepare_visdrone.py --root datasets/VisDrone
python scripts/analyze_small_objects.py --labels-dir datasets/VisDrone/labels/train --output results/visdrone_small_object_stats.json

The conversion keeps VisDrone boxes with score == 1 and object_category in 1..10, then remaps categories to zero-based YOLO labels.

Run experiments

Quick static argument check without loading Ultralytics:

python scripts/run_train.py --variant baseline --data configs/visdrone.yaml --dry-run

Single variant training:

DATA=configs/visdrone.yaml IMGSZ=960 EPOCHS=50 BATCH=8 DEVICE=0 bash scripts/train.sh baseline
DATA=configs/visdrone.yaml IMGSZ=960 EPOCHS=50 BATCH=4 DEVICE=0 bash scripts/train.sh p2

Sequential ablation:

DATA=configs/visdrone.yaml IMGSZ=960 EPOCHS=50 bash scripts/ablation.sh

Official-checkpoint public recipe approximation:

DRY_RUN=1 DATA=configs/coco2017.example.yaml bash scripts/run_official_yolo26n_recipe_approx.sh

Results

See RESULTS.md for public-safe metrics and the official-checkpoint reproduction boundary.

License

AGPL-3.0-only. This project uses Ultralytics YOLO tooling, which is available under AGPL-3.0 or a commercial Ultralytics Enterprise license.