khanhromvn/waste-detection-yolo11

Waste Detection YOLO11 (organic / recyclable / residual)

Object detection model that takes ONE photo and returns MANY bounding boxes, each labelled with one of 3 waste classes. Fine-tuned from yolo11m.pt (Ultralytics YOLO11).

• Author: KhanhRomVN (khanhromvn@gmail.com)
• Model file: waste_detection_yolo11m.pt (38.6 MB, 20.06M params)
• Framework: Ultralytics 8.4.182, PyTorch 2.11.0
• Task: object detection (not whole-image classification)
• License: MIT (check the licenses of the original datasets before commercial use)

Classes

id	name	examples
0	organic	food scraps, fruit, vegetables, bread, pizza, eggshell
1	recyclable	plastic bottle, drink can, tin can, glass jar/bottle, cardboard, paper, metal lid
2	residual	plastic bag/film, wrapper, crisp packet, styrofoam, disposable cup/cutlery, tissue, battery, cigarette butt, broken glass

Demo: model output on real images

Left: original image. Right: model prediction (bounding boxes with class and confidence).

Original	Result

More pairs: images/demo/ (8 pairs, N_original.png / N_result.png).

Quick start

pip install ultralytics huggingface_hub

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

path = hf_hub_download(repo_id="khanhromvn/waste_classification", filename="waste_detection_yolo11m.pt")
model = YOLO(path)
results = model.predict("your_image.jpg", conf=0.25, imgsz=640)
results[0].save("result.jpg")
for b in results[0].boxes:
    print(model.names[int(b.cls)], float(b.conf), b.xyxy.tolist())

Or use the CLI script in this repo:

pip install -r requirements.txt
python predict.py --source your_image.jpg --conf 0.25

Or the Ultralytics CLI with a local file:

yolo detect predict model=waste_detection_yolo11m.pt source=your_image.jpg conf=0.25

Recommended confidence threshold: 0.25 (see images/yolo11m/BoxF1_curve.png to pick your own).

Results (VAL: 441 images, 1,159 boxes, conf 0.001, IoU 0.6)

The TEST split was never used for model selection or reported here.

Model	Params (M)	File size	Epochs run	Best epoch	mAP50	mAP50-95	P	R	ms/img (T4)	Train min
yolo11m (this model)	20.06	38.6 MB	32	24	0.495	0.388	0.606	0.463	24.0	241
yolo11s	9.43	18.3 MB	54	34	0.481	0.373	0.602	0.450	8.5	206
yolo11n	2.61	5.2 MB	55	40	0.455	0.349	0.541	0.453	3.8	181

This repo ships yolo11m (best accuracy). The yolo11s and yolo11n numbers are listed for comparison; their charts are in images/.

Per class AP50 / AP50-95 (yolo11m, VAL)

Class	Boxes	AP50	AP50-95	P	R
organic	396	0.519	0.398	0.606	0.446
recyclable	462	0.515	0.407	0.671	0.465
residual	301	0.450	0.357	0.540	0.478

Per source (yolo11m, VAL)

Source	Images	mAP50	mAP50-95	Note
oi (Open Images v7)	160	0.600	0.532	labels incomplete, only 3 residual boxes
taco	281	0.345	0.272	only 6 organic boxes

Charts

Demo pairs (original vs result) are in images/demo/ (N_original.png / N_result.png).

Training data

12,602 images (train 11,771 / val 441 / test 390), 9 sources mapped to 3 classes. Boxes: train 19,986 / val 1,159 / test 1,096. Train boxes per class: organic 6,119, recyclable 7,842, residual 6,025.

Source	Images	Box type
fruits360	295	auto-generated
g10v2 (10-class garbage)	1,420	auto-generated
g12 (12-class garbage)	2,139	auto-generated
oi (Open Images v7)	1,052	real boxes
realwaste	2,240	auto-generated
synth (synthetic composites)	1,500	composite
taco	1,390	real boxes
techsash (O/R set)	1,779	auto-generated
trashnet	787	auto-generated

Val and test come only from the two sources with real boxes (taco, oi). Real-box training images are oversampled x2. Preprocessing: deduplication, CLIP zero-shot check of noisy labels (plastic bags moved from recyclable to residual), auto-generated boxes for label-only sources, synthetic composites. Label mapping per source: results/label_map.csv.

Training settings

imgsz 640, optimizer auto, lr0 0.01, lrf 0.01, warmup 3 epochs, weight decay 0.0005, mosaic 1.0 (off for last 10 epochs), mixup 0.1, HSV 0.015/0.7/0.4, fliplr 0.5, rotation 5 deg, translate 0.1, scale 0.5, AMP on, deterministic, seed 42. Hardware: Google Colab Free, Tesla T4. yolo11m: batch 16, patience 20, 4 h cap (stopped at epoch 32). Full details and scripts: training/README.md.

Limitations

• Moderate accuracy (mAP50 0.495); the main cause is data quality, not model size.
• Open Images annotations are incomplete, so precision/mAP on oi are artificially low.
• Rare class/source combinations (taco organic: 6 boxes, oi residual: 3 boxes) make some per-class numbers unreliable.
• Most training boxes are auto-generated, only taco and oi have real boxes.
• yolo11m was cut by a time cap at epoch 32.
• TEST split (390 images, 1,096 boxes) has not been evaluated yet.
• residual is the hardest class. Not trained on clothes, shoes, textiles, electronics, light bulbs.
• Not for safety-critical use.

Repo layout

waste_detection_yolo11m.pt   model weights (Ultralytics .pt)
predict.py                   inference script
requirements.txt
images/                      curves, confusion matrices, demo images (yolo11m / yolo11s / yolo11n)
results/                     metrics CSV, manifest.json, dataset statistics, label_map.csv, classes.json
training/                    phase1-4 scripts, args, config, logs

SHA-256 of waste_detection_yolo11m.pt: 9f2d19a7ce386c6dbb95a725e6197f7b6fb699cd601bbeca9407034396f2ad6a

Tom tat (Tieng Viet)

Mo hinh YOLO11m phat hien va phan loai rac thai trong anh thanh 3 nhom: huu co (organic), tai che (recyclable), rac con lai (residual). Huan luyen tren 12.602 anh tu 9 nguon du lieu, mAP50 = 0.495 va mAP50-95 = 0.388 tren tap val. Cach chay: xem phan Quick start o tren.

Citation

@misc{khanhromvn2026wastedetection,
  author = {KhanhRomVN},
  title = {Waste Detection YOLO11 (organic / recyclable / residual)},
  year = {2026},
  publisher = {Hugging Face},
  email = {khanhromvn@gmail.com}
}

Built with Ultralytics YOLO11.