MobileViT-S Garbage Classifier

Binary classification model for filtering objectively-bad frames (black, blurry, uniform-color, low-detail) in anime video preprocessing pipelines.

Model Details

• Architecture: MobileViT-S
• Parameters: 4.94M
• Model Size: 20MB
• Input Size: 256×256
• Classes: [quality, garbage]

Performance

Without threshold (0.5):

• Accuracy: 93.46%
• Precision: 92.24%
• Recall: 95.47%
• F1-Score: 93.83%

With optimal threshold (0.7115):

• Accuracy: 93.62%
• Precision: 93.92%
• Recall: 93.82%
• F1-Score: 93.87%

Usage

import torch
import timm
from torchvision import transforms
from PIL import Image

# Load model
model = timm.create_model('mobilevit_s', num_classes=2, pretrained=False)
model.load_state_dict(torch.load('pytorch_model.bin'))
model.eval()

# Prepare image
transform = transforms.Compose([
    transforms.Resize((256, 256)),
    transforms.ToTensor(),
    transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
])

img = transform(Image.open('frame.webp').convert('RGB')).unsqueeze(0).cuda()

# Predict
with torch.no_grad():
    logits = model(img)
    probs = torch.softmax(logits, dim=1)
    garbage_prob = probs[0, 0].item()  # Class 0 = garbage

# Decision
is_garbage = garbage_prob > 0.7115  # Use optimal threshold

Training Data

• Total frames: 12,440
• Training: 10,574 frames
• Validation: 1,866 frames (895 garbage, 971 quality)
• Labeling: Verified via reverse-engineered frame matching

Garbage Detection

Filters frames with:

• Solid black/white/uniform color (33%)
• No edge patterns (33%)
• Low detail content (16%)
• Extreme outliers (15%)

Threshold Recommendations

• Default (0.5): Good starting point, slightly higher recall
• Optimal (0.7115): Best F1-score, balanced precision/recall
• High precision (0.75-0.80): Reduce false positives
• High recall (0.60-0.65): Catch more garbage, accept more false positives

License

MIT