Apartment Room Classifier

• Fine-tuned ViT model Scampolonii/vit-apartment-rooms
• Zero-shot CLIP (openai/clip-vit-large-patch14)
• OpenAI vision model (GPT-4o image classification)

Dataset Used For Training

• Hugging Face dataset: keremberke/indoor-scene-classification (MIT Indoor Scenes, repackaged by Roboflow)
• Original source: MIT Indoor Scene Recognition
• Total images: ~15,571 (train: 10,885 / val: 3,128 / test: 1,558)
• Number of classes used: 8 (filtered from 67 total MIT classes)

Trained Model

• Hugging Face model link: https://huggingface.co/Scampolonii/vit-apartment-rooms
• Base model: google/vit-base-patch16-224
• Training strategy: Transfer learning — all ViT layers frozen, only classifier head trained (4,614 trainable parameters out of 85,803,270)

Preprocessing Steps

• Filtered MIT Indoor Scenes to 8 apartment-relevant room categories
• Remapped labels to new 0-based IDs
• Split dataset: 80% train / 10% validation / 10% test
• All images converted to RGB
• Resized to 224×224 pixels (via ViTImageProcessor)
• Normalized with mean=[0.5, 0.5, 0.5] and std=[0.5, 0.5, 0.5]
• Rescale factor: 1/255

Training Performance

Per-epoch metrics were not logged (report_to="none"). Final test accuracy: 90.16% on the held-out test set.

Evaluation

• Metric: Accuracy
• Evaluated using Hugging Face evaluate library
• Test set accuracy: 90.16% (eval_loss: 0.3788)

Example Image Results

Model Comparison Summary

Links

• Live App (HF Space): https://huggingface.co/spaces/Scampolonii/apartment-room-classifier
• Model: https://huggingface.co/Scampolonii/vit-apartment-rooms