biaeye-yolo11s

BiaEye litter detector — YOLO11s fine-tuned on TACO (six materials)

Detector behind the BiaEye platform (AI illegal-waste detection for Amman: citizen app, truck/edge cameras, operator dashboard). Ultralytics YOLO11s fine-tuned from the COCO checkpoint on the TACO street-litter dataset, with TACO's 60 categories grouped into the six materials the platform uses:

0 plastic · 1 paper · 2 glass · 3 metal · 4 organic · 5 bulky

Training

• 1,500 TACO images (long side resized to 1280 px), split 75 / 15 / 10 → 1,125 train, 225 val, 150 test
• yolo11s.pt, 80 epochs, 640 px, batch 8, AdamW-style defaults with cosine LR, mosaic / flip / HSV / rotation augmentation, dropout 0.15
• RTX 3050 Laptop (4 GB), 42 minutes

Results on the held-out 150 test images (never seen in training)

TACO is small and imbalanced (3,437 plastic boxes vs 8 organic), so the per-material scores are modest; organic and bulky are effectively unlearned. Detection of litter lying on the ground is the useful part: on phone photos it finds a crumpled tissue on a floor at 0.83–0.88 confidence and ignores bins and furniture.

Use

from huggingface_hub import hf_hub_download
from ultralytics import YOLO
model = YOLO(hf_hub_download("raneemsadeh/biaeye-yolo11s", "best.pt"))
model.predict("street.jpg", conf=0.35, imgsz=640)

In BiaEye: MODEL_WEIGHTS=hf://raneemsadeh/biaeye-yolo11s/best.pt.

Files

best.pt weights · results.png training curves · BoxPR_curve.png · confusion_matrix_normalized.png · args.yaml · results.csv