Scene Recognition

Classification of 365 Scenes by fine-tuning Vision Transformer on Places365 Standard dataset.

Example Scenes

The Places365 dataset contains 365 Scenes. Below are a few example scenes.

Images	Labels
	Baseball Field
	Balcony Interior
	Embassy
	Fire Escape
	Kitchen
	Lake Natural
	Skyscraper
	Office Cubicles
	Reception

Training

The Vision Transformer model was finetuned on the Places365 dataset with the hyperparmeters as follows:

Hyperparameter	Value
Batch Size	32
Learning Rate	2e-4
Optimizer	AdamW
No of Epochs	5
Evaluate Validation after	5000 batches

The metrics were logged with the help of Weights and Biases. This specific run can be found here.

Evaluation on the Test dataset

The Trained Vision Transformer Model was evaluated on the Places365 test dataset and obtained the following results:

Metric	Value
AUROC	98.90
Accuracy Top 5	83.52
Accuracy Top 1	52.47
F1-Score	51.71
Precision	52.70
Recall	52.47

How to Run

1. Download the Places365 Standard dataset.
2. Install the requirements from requirements.txt.
3. Update the paths of
   • DATASET_TRAIN_PATH : Path of Places365 Standard training dataset.
   • DATASET_TEST_PATH : Path of Places365 Standard validation dataset.
   • DATASET_MAPPINGS_PATH : Path to store the dataset mappings for train and test datasets.
   • WANDB_PATH : Path to initialize Weights and Bias runs.
4. Run preprocess_dataset.py to create a mapping of images.
5. Train the model by running the train.py script.
6. Evaluate the model on the test dataset by running the evaluate_test.py script.