rishi506/clip-car-damage

CLIP Fine-tuned: Automotive Damage Detection

Fine-tuned openai/clip-vit-base-patch32 with LoRA adapters for classifying vehicle damage from customer photos — insurance claim use case.

Classes

• dent: a car with a dent or body panel deformation
• intact: an intact car with no visible damage
• major_collision: a severely damaged car after a major collision
• scratch: a car with paint scratches or surface abrasions

Performance

• Test Accuracy: 69.79%
• Training: 15 epochs, InfoNCE contrastive loss
• LoRA rank: 16 (only 0.0M trainable params)

Usage

import torch

model     = CLIPModel.from_pretrained("rishi506/clip-car-damage")
processor = CLIPProcessor.from_pretrained("rishi506/clip-car-damage")

# Zero-shot classification
texts = list(['a car with a dent or body panel deformation', 'an intact car with no visible damage'] ...)
inputs = processor(text=texts, images=your_image, return_tensors="pt", padding=True)
outputs = model(**inputs)
probs = outputs.logits_per_image.softmax(dim=1)