W4ashabii/waste_classifier

Waste Classifier (Bio vs Non-Bio) — YOLOv8s-cls

A lightweight image classifier that sorts waste photos into biodegradable (Bio) and non-biodegradable (Non_Bio). It is YOLOv8s-cls fine-tuned from the COCO/ImageNet-pretrained checkpoint on the Waste_SIP_Dataset.

• Task: binary image classification
• Classes: Bio, Non_Bio
• Architecture: YOLOv8s-cls (~10 MB checkpoint)
• Input size: 224 × 224
• Best validation top-1 accuracy: 0.96 (epoch 19 of 20)

Quick start

pip install ultralytics huggingface_hub

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

weights = hf_hub_download(repo_id="W4ashabii/waste_classifier", filename="best.pt")
model = YOLO(weights)

results = model.predict("your_image.jpg", imgsz=224)
r = results[0]
print(r.names[r.probs.top1], float(r.probs.top1conf))

Batch prediction on a folder:

results = model.predict("path/to/images/", imgsz=224)
for r in results:
    print(r.path, r.names[r.probs.top1], round(float(r.probs.top1conf), 3))

Results

Metrics from the validation split, using the best checkpoint (best.pt, epoch 19):

Metric	Value
Top-1 accuracy	0.960
Final-epoch (20) top-1 accuracy	0.955
Final training loss	0.038

Per-class recall (from the normalized confusion matrix):

Class	Recall
Bio	0.97
Non_Bio	0.95

Top-5 accuracy is always 1.0 because there are only two classes, so it is not informative.

Training lasted about 2.3 minutes (140 s for 20 epochs).

Note: best.pt was chosen by validation accuracy, so the validation numbers above are slightly optimistic. For an unbiased estimate, evaluate on the dataset's held-out test split.

Training details

Setting	Value
Base model	yolov8s-cls.pt (pretrained)
Framework	Ultralytics YOLOv8
Epochs	20 (early-stopping patience 5)
Image size	224
Batch size	64
Optimizer	AdamW
Initial learning rate	1e-3, cosine schedule
Precision	FP16 mixed precision (AMP)
Hardware	NVIDIA RTX 4070 (8 GB)

Dataset

Trained on Pramudit/Waste_SIP_Dataset, an image-folder dataset with train, validation and test splits and two classes (Bio, Non_Bio). Please refer to the dataset page for its license, collection method and class distribution.

Intended use

• Demos, teaching and prototyping of automated waste sorting
• A baseline for waste-classification research
• Starting point for further fine-tuning on your own waste images

Limitations

• Binary only. The model tells biodegradable from non-biodegradable. It does not identify material types (plastic, glass, metal, paper) and does not detect or locate objects in an image.
• Single-object assumption. It classifies the whole image, so photos with mixed waste or cluttered scenes may give unreliable results.
• Domain shift. Accuracy was measured on images from one dataset. Performance can drop with different cameras, lighting, backgrounds, or waste types that are rare in the training data.
• Not safety-critical. Do not rely on it as the sole decision-maker for real waste handling, hygiene or compliance.

License

The weights are released under AGPL-3.0, in line with the Ultralytics YOLOv8 license. The training data has its own license; check the dataset page before commercial use.

Citation

@software{jocher2023ultralytics,
  author  = {Jocher, Glenn and Chaurasia, Ayush and Qiu, Jing},
  title   = {Ultralytics YOLOv8},
  version = {8.0.0},
  year    = {2023},
  url     = {https://github.com/ultralytics/ultralytics},
  license = {AGPL-3.0}
}