smart-waste-segregator

Smart Waste Segregator: model weights

Weights for the Smart Waste Segregator, a Raspberry Pi bin that classifies incoming trash and sorts it with a dual-flap servo mechanism.

Code, notebooks and evaluation artifacts: https://github.com/divyanshuklai/smart-waste-segregator

Contents

mobilenetv3-small/
  best_model.pth          18.5 MB   trained PyTorch checkpoint
  waste_classifier.onnx    6.1 MB   ONNX export, this is what runs on the Pi
yolo11/
  yolo11n.pt               5.6 MB   detection weights
  yolo11n-cls.pt           5.8 MB   classification weights
  yolo11n.torchscript     11.1 MB   TorchScript export
  yolo11n-cls.torchscript 11.5 MB   TorchScript export
  yolo11n_ncnn_model/               NCNN export (model.ncnn.bin, .param, model_ncnn.py, metadata.yaml)
  yolo11n-cls_ncnn_model/           NCNN export

MobileNetV3-Small classifier

The shipped model is MobileNetV3-Small, not MobileNetV2. A MobileNetV2 was trained as a separate experiment but did not ship. The checkpoint here has squeeze-and-excite blocks and the ONNX graph contains 28 HardSigmoid nodes, both of which identify it as V3-Small.

Classes, in the order the shipped weights expect:

0  metal
1  plastic
2  glass
3  biodegradable

Training

A second run in the same notebook used 20 epochs, 165 batches per epoch, OneCycleLR, early stopping with patience 5, and heavier augmentation. Per-epoch numbers were not saved.

ONNX export

On-device inference uses onnxruntime with CPUExecutionProvider, GraphOptimizationLevel.ORT_ENABLE_ALL and intra_op_num_threads = 4. Images are captured with libcamera-still at 2592x1944, converted BGR to RGB with OpenCV and resized to 224x224.

import onnxruntime, numpy as np

opts = onnxruntime.SessionOptions()
opts.graph_optimization_level = onnxruntime.GraphOptimizationLevel.ORT_ENABLE_ALL
opts.intra_op_num_threads = 4

sess = onnxruntime.InferenceSession(
    "mobilenetv3-small/waste_classifier.onnx",
    providers=["CPUExecutionProvider"],
    sess_options=opts,
)

x = np.random.rand(1, 3, 224, 224).astype(np.float32)
logits = sess.run(None, {sess.get_inputs()[0].name: x})[0]
print(["metal", "plastic", "glass", "biodegradable"][int(logits.argmax())])

YOLO11 edge exports

A later pass exporting YOLO11n to formats that run on constrained hardware.

NCNN classifier timing on the bus.jpg sample: 124.0 ms inference, 78.8 ms preprocess, 1.3 ms postprocess.

Limitations

Read these before using the weights for anything.

• Real-world performance is poor. The training data came from public waste datasets whose images do not resemble real waste in a real bin.
• There is no valid test number. The 91.63% is validation accuracy measured on the same biased distribution the model trained on.
• The test notebook's 2.75% accuracy is a bug, not a result. It reads raw indices from a six-class YOLO label set (BIODEGRADABLE, CARDBOARD, GLASS, METAL, PAPER, PLASTIC), drops every index >= 4, which discards 860 of 1,042 images, then reads the surviving indices against a different four-class list. The large off-diagonal count in the repository's confusion matrix is that mismapping, not model error.
• Three class orderings exist across the source repository. The order given above is the one the shipped weights expect. The Pi runtime script smart_waste_segregation.py uses ['plastic', 'paper', 'metal', 'glass'], which is wrong for the model it loads, so the device printed mislabelled class names. This is documented rather than silently corrected.
• The YOLO11 weights are stock Ultralytics checkpoints and their exports. They were not fine-tuned on waste data.

Citation

Divyansh Shukla, 2024-2025. Smart Waste Segregator.