Ank0805/waste-classification

YOLO11n Granular Waste Material Classifier (42 Classes)

Model Summary

This repository provides an optimized YOLO11 nano-architecture object detection model trained for granular classification and spatial localization of solid municipal refuse across 42 discrete material categories.

The model is specialized in distinguishing subtle visual differences among polymers, cellulosic fibers, ferrous and non-ferrous metals, mineral glass, and composite packaging. Operating at a native resolution of 640x640 with 2.59 million parameters and 6.5 GFLOPs, it is designed for edge and embedded robotic sorting applications, municipal collection analysis, and high-throughput conveyor inspection.

Model Details

• Architecture: Ultralytics YOLO11n (Decoupled anchor-free detection head, C3k2, SPPF, C2PSA)
• Task: 2D Bounding Box Object Detection and Multi-Class Material Categorization
• Input Dimensions: 640 x 640 x 3 (RGB)
• Parameters: 2,590,425
• GFLOPs: 6.5
• Framework: PyTorch / Ultralytics

Granular Material Taxonomy (42 Classes)

The model discriminates among the following 42 material classifications:

Class ID	Class Name	Material Category
0	Aerosols	Pressurized metallic spray canisters
1	Aluminum can	Non-ferrous aluminum beverage containers
2	Aluminum caps	Stamped aluminum closures and twist tops
3	Cardboard	Corrugated and rigid fiberboard packaging
4	Cellulose	Pure cellulosic fiber and cellulose-derived packaging
5	Ceramic	Vitrified ceramics, porcelain, and clay fragments
6	Combined plastic	Multi-layer or composite polymer structures
7	Container for household chemicals	High-Density Polyethylene (HDPE) detergent/chemical bottles
8	Disposable tableware	Single-use plastic and paper cutlery, plates, and trays
9	Electronics	Circuit boards, e-waste, wiring, small electronic appliances
10	Foil	Thin aluminum foil and metallic wrapping sheets
11	Furniture	Discarded domestic and industrial composite furniture
12	Glass bottle	Molded soda-lime glass bottles and jars
13	Iron utensils	Ferrous cookware, cutlery, and kitchen implements
14	Liquid	Visible fluid pools, spilled viscous liquids
15	Metal shavings	Industrial turnings, swarf, and metal particulate waste
16	Milk bottle	Translucent HDPE or PET dairy containers
17	Organic	Food scraps, agricultural biomass, decomposed matter
18	Paper bag	Kraft and bleached cellulosic carrier bags
19	Paper cups	Polyethylene-coated cellulosic beverage cups
20	Paper shavings	Shredded paper and industrial cellulosic off-cuts
21	Paper	Standard printed, bleached, or packaging paper sheets
22	Papier mache	Molded pulp packaging (egg cartons, protective cushions)
23	Plastic bag	Low-Density Polyethylene (LDPE) flexible shopping bags
24	Plastic bottle	Polyethylene Terephthalate (PET) beverage bottles
25	Plastic can	Rigid molded polymer cans and cylindrical packaging
26	Plastic canister	Heavy-duty HDPE industrial storage canisters
27	Plastic caps	Polypropylene (PP) and HDPE bottle closures
28	Plastic cup	Polypropylene (PP) and Polystyrene (PS) drink cups
29	Plastic shaker	Molded polymer cosmetic and beverage dispensers
30	Plastic shavings	Shredded plastic flake and micro-particulate waste
31	Plastic toys	Injection-molded ABS, PVC, and PE structural toys
32	Postal packaging	Padded mailers, bubble wrap, poly mailer bags
33	Printing industry	Coated card stock, offset printed material
34	Scrap metal	Structural ferrous and non-ferrous construction scrap
35	Stretch film	Linear Low-Density Polyethylene (LLDPE) pallet wrap
36	Tetra pack	Aseptic liquid cartons (paperboard/foil/polyethylene laminate)
37	Textile	Woven and non-woven fabrics, clothing, industrial rags
38	Tin	Tinplate food cans and preserve containers
39	Unknown plastic	Unspecified or degraded structural polymer fragments
40	Wood	Lumber, pallets, dimensional timber fragments
41	Zip plastic bag	Reusable LDPE slide and press-to-close storage bags

Quickstart and Usage

Prerequisites

pip install ultralytics opencv-python huggingface_hub torch torchvision

Method 1: Automated Hugging Face Hub Execution (For Evaluators)

Evaluators can run inference directly by pulling weights from the Hugging Face repository without manual file downloads:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO
import cv2

# Pull model weights directly from repository
model_path = hf_hub_download(
    repo_id="Ank0805/waste-classification",  # Replace with exact repo path if different
    filename="waste classification modle.pt"
)

# Initialize model
model = YOLO(model_path)

# Execute inference
results = model.predict(
    source="path/to/test_image.jpg",
    conf=0.25,
    iou=0.45,
    imgsz=640,
    device=0  # Set to 'cpu' if no NVIDIA GPU is available
)

# Parse output detections
for result in results:
    for box in result.boxes:
        cls_id = int(box.cls[0].item())
        confidence = float(box.conf[0].item())
        bbox = box.xyxy[0].tolist()
        label = model.names[cls_id]
        print(f"Material: {label:<30} | Confidence: {confidence:.3f} | BBox: {bbox}")

    # Save visual rendering
    annotated_frame = result.plot()
    cv2.imwrite("classification_output.jpg", annotated_frame)

Method 2: Local Python API

import cv2
from ultralytics import YOLO

# Load local weights
model = YOLO("waste classification modle.pt")

# Execute inference
results = model.predict(
    source="path/to/image.jpg",
    conf=0.25,
    iou=0.45,
    imgsz=640,
    device=0
)

# Parse output detections
for result in results:
    for box in result.boxes:
        cls_id = int(box.cls[0].item())
        label = model.names[cls_id]
        conf = float(box.conf[0].item())
        print(f"Detected: {label} ({conf:.2f})")

    annotated_frame = result.plot()
    cv2.imwrite("output.jpg", annotated_frame)

Method 3: Video Stream and Conveyor Inspection

from ultralytics import YOLO

model = YOLO("waste classification modle.pt")

# Stream inference from video file, USB webcam (0), or industrial RTSP stream
results = model.predict(
    source="conveyor_stream.mp4",
    conf=0.25,
    iou=0.45,
    save=True,
    stream=True
)

for frame_result in results:
    pass  # Generator executes per-frame inference

Method 4: Command-Line Interface (CLI)

# Execute prediction on an image
yolo task=detect mode=predict model="waste classification modle.pt" source="image.jpg" conf=0.25 save=True

# Execute prediction on a video directory using GPU
yolo task=detect mode=predict model="waste classification modle.pt" source="video_dir/" device=0 conf=0.25 save=True

Method 5: Compilation and Edge Engine Export

from ultralytics import YOLO

model = YOLO("waste classification modle.pt")

# Export to ONNX for cross-platform CPU/GPU deployment
model.export(format="onnx", imgsz=640, simplify=True)

# Export to NVIDIA TensorRT for ultra-low latency FP16 inference
model.export(format="engine", imgsz=640, half=True, device=0)

# Export to Intel OpenVINO
model.export(format="openvino", imgsz=640, half=True)

# Export to Tencent NCNN for embedded ARM hardware (Raspberry Pi, Jetson)
model.export(format="ncnn", imgsz=640)

Datasets and Data Sources

• ProjectVerba / YoloTrash Benchmark: Roboflow Universe Repository
  Annotated ground-level and close-range multi-class refuse corpus spanning 42 discrete material categories.

Training Specifications

• Optimization Algorithm: AdamW ($\beta_1 = 0.937, \beta_2 = 0.999$, weight decay = $5 \times 10^{-4}$)
• Epochs: 120
• Learning Rate Schedule: Initial $\eta_0 = 1\times 10^{-3}$, final factor $\text{lrf} = 0.01$ via Cosine Annealing
• Terminal Augmentation Deactivation: Mosaic augmentation active for epochs 1 to 100, fully deactivated (close_mosaic=20) during epochs 101 to 120 to enable convergence on natural, unmanipulated object silhouettes.

Operational Recommendations

1. Recycling Facility / High-Recall Mode: For initial sorting passes where omitting recyclable material carries a high cost, operate at conf=0.20 to conf=0.25.
2. Robotic Pick-and-Place Mode: For automated robotic arms requiring strict material certainty before physical grasping, set conf >= 0.45 to prevent picking errors on visually degraded items.