DINOv2 Garbage Classification

A fine-tuned DINOv2 ViT-L/14 model for garbage image classification.

This repository is prepared as a workshop baseline model, allowing participants to quickly start experimenting with transfer learning techniques without training a Vision Transformer from scratch.

The checkpoint can be used directly for inference or as an initialization for further fine-tuning on custom datasets.

Overview

This project fine-tunes Meta AI's DINOv2 ViT-L/14 backbone for a three-class garbage classification task.

Unlike conventional CNN-based approaches, DINOv2 learns powerful visual representations through self-supervised learning, making it an excellent backbone for downstream computer vision tasks with limited labeled data.

Participants are encouraged to extend this model with techniques such as:

• Hard Example Mining
• Active Learning
• Semi-supervised Learning
• Domain Adaptation
• Knowledge Distillation
• Custom Classification Heads
• Full Backbone Fine-tuning
• Test-Time Augmentation (TTA)

Model Architecture

Input Image (518 × 518)
        │
        ▼
DINOv2 ViT-L/14 Backbone
        │
        ▼
1024-D Feature Vector
        │
        ▼
LayerNorm
        │
        ▼
Linear (1024 → 512)
        │
        ▼
GELU
        │
        ▼
Dropout (0.3)
        │
        ▼
Linear (512 → 3)
        │
        ▼
Prediction

Backbone

• Base Model: facebookresearch/dinov2
• Variant: dinov2_vitl14
• Feature Dimension: 1024
• Framework: PyTorch

Training strategy:

• Freeze entire backbone
• Unfreeze last 4 transformer blocks
• Train custom classification head

Dataset

The model classifies garbage into three categories.

Sample Images

Recyclable

Electronic

Organic

Dataset Statistics

Class Distribution

Data Augmentation

Training images are augmented using:

• Random Resized Crop
• Horizontal Flip
• Color Jitter
• Gaussian Blur
• RandAugment
• ImageNet Normalization

Input resolution:

518 × 518

Repository Structure

.
├── assets/
│   ├── 0_Recyclable/
│   ├── 1_Electronic/
│   └── 2_Organic/
│
├── best_dinov2.pth
└── README.md

Load the Model

import torch
import torch.nn as nn
from huggingface_hub import hf_hub_download

pth_path = hf_hub_download(
    repo_id="galihkjaya/DINOv2-Garbage-Classification",
    filename="best_dinov2.pth"
)

class DINOv2Classifier(nn.Module):
    def __init__(self, num_classes=3, unfreeze_blocks=4):
        super().__init__()

        self.backbone = torch.hub.load(
            "facebookresearch/dinov2",
            "dinov2_vitl14"
        )

        for p in self.backbone.parameters():
            p.requires_grad = False

        for block in self.backbone.blocks[-unfreeze_blocks:]:
            for p in block.parameters():
                p.requires_grad = True

        self.head = nn.Sequential(
            nn.LayerNorm(1024),
            nn.Linear(1024, 512),
            nn.GELU(),
            nn.Dropout(0.3),
            nn.Linear(512, num_classes)
        )

    def forward(self, x):
        features = self.backbone(x)
        return self.head(features)

model = DINOv2Classifier()

model.load_state_dict(torch.load(pth_path, map_location="cpu"))
model.eval()

Inference Example

from PIL import Image
from torchvision import transforms

transform = transforms.Compose([
    transforms.Resize((518, 518)),
    transforms.ToTensor(),
    transforms.Normalize(
        mean=(0.485,0.456,0.406),
        std=(0.229,0.224,0.225)
    )
])

labels = {
    0: "Recyclable",
    1: "Electronic",
    2: "Organic"
}

image = Image.open("sample.jpg").convert("RGB")
image = transform(image).unsqueeze(0)

with torch.no_grad():
    logits = model(image)
    pred = logits.argmax(1).item()

print(labels[pred])

Future Improvement

This checkpoint serves as the baseline model. Suggested follow-up experiments include:

• Fine-tune all transformer blocks
• Hard Example Mining
• Replace the classification head
• Add new waste categories
• Train with Focal Loss
• Label Smoothing
• Test-Time Augmentation
• MixUp / CutMix
• Semi-supervised Learning
• Domain Adaptation
• Feature Extraction using DINOv2 embeddings

The objective is to demonstrate how a strong pretrained visual backbone can be adapted efficiently for downstream classification tasks.

Limitations

• Only supports three waste categories.
• Performance depends on image quality and lighting conditions.
• Mixed-material waste may be difficult to classify.
• Intended for educational and research purposes.

Acknowledgements

This project builds upon:

• Meta AI — DINOv2
• PyTorch
• Hugging Face Hub
• Kaggle

Citation

@misc{galih2026,
  title={DINOv2 Garbage Classification},
  author={Galih Kusuma Wijaya},
  year={2026},
  publisher={Hugging Face},
  howpublished={https://huggingface.co/galihkjaya/DINOv2-Garbage-Classification}
}