foodseg103-deeplabv3plus-resnet50

FoodSeg103 Semantic Segmentation - DeepLabV3+ (ResNet50)

A semantic segmentation model trained on the FoodSeg103 dataset for food image segmentation.

Model Architecture

• Architecture: DeepLabV3+
• Backbone: ResNet50 (ImageNet pretrained)
• Framework: PyTorch + segmentation_models_pytorch
• Classes: 104 (103 food categories + background)

Training Details

• Epochs: 44
• Best Validation IoU: 0.6686 (Epoch 30)
• Final Validation IoU: 0.6533
• Image Size: 256x256
• Optimizer: Adam with ReduceLROnPlateau scheduler
• Loss: Cross Entropy

Training Curves

Loss Curves

Validation IoU

Training Summary

Usage

import torch
import segmentation_models_pytorch as smp
from PIL import Image
import torchvision.transforms as T

# Load model
model = smp.DeepLabV3Plus(
    encoder_name="resnet50",
    encoder_weights=None,
    in_channels=3,
    classes=104,
)
model.load_state_dict(torch.load("best_model.pth", map_location="cpu"))
model.eval()

# Preprocess image
transform = T.Compose([
    T.Resize((256, 256)),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225])
])

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

# Inference
with torch.no_grad():
    output = model(input_tensor)
    pred_mask = output.argmax(dim=1).squeeze().numpy()

Dataset

Trained on FoodSeg103 - a large-scale benchmark for food image segmentation with 103 food categories.

License

MIT