A deep learning model which can recognize a scene of 10 different classes:
- inside an airport
- Bakery
- Bedroom
- Greenhouse
- Gym
- Kitchen
- Operating room
- Pool
- Restaurant
- Toystore
The model achieved an accuracy of 94.5% on the test data by finetuning an inception-resnet-v2 model. This was a competition on Kaggle which won first place: https://www.kaggle.com/c/fcis-cs-deeplearningcompetition/leaderboard, Team14.
The project was run on google colab but if you want to run on your own device you will need to install the following:
- cv2
- numpy
- tensorflow
- matplotlib
- glob
- PIL
- tqdm
The dataset that is available on the competition was converted into tfrecord to be easier to use in training. The training data was split into training and validation sets, which can be found here. The testing data can be found here.
The training & validation data should be placed in folder: train_data and the testing data should be placed in folder: test_data.
The model used to achieve the accuracy can be found here.
The files of the model should be placed in folder: models.
- To train your own model you should use the file: training_inception_resnet_v2.py
- To run the model on the validation set use the file: validation_file.py
- To run the model on the testing data use the file: testing_file.py