Scene classification is a challenging but important image classification task. Variation in direction, weather, time and subject recognition directly complicate the process. Historically, deep learning is rapidly overtaking classical approaches for scene classification. However, the deep learning approach needs a large number of images. Zhou et al. shows that the Alexnet and successful deep learning approach in general image classification actually achieve similar performance with large scene image dataest. In this project, we does scene classification by only using a small amount of images.
The data used here is from Miniplaces Challenge. The goal of that challenge is to identify the scene category depicted in a photograph. The origin data is from the Places2 dataset, which has 10+ million images and 400+ unique scene categories. The data used in this project only has 100 scene categories and 100,000 images for training, 9,000 images for validation, and 1,000 images for testing.
We divide all data into three dataset.
Number of images
Dataset TRAIN VALIDATION TEST
------------------------------------------------
MiniPlaces 100,000 9,000 1,000Database statistics is as follows:
Training: 100,000 images, with 1000 images per category
Validation: 10,000 images, with 100 images per category
Test: 10,000 images, with 100 images per categoryconda install numpyconda install scipyconda install tensorflowFor GPU user:
conda install tensorflow-gpuconda install matplotlibconda install -c menpo opencvWe already packaged all the data you need to download to a script. Simple run follwoing command to get all data.
bash get_data.shIn order to run the network, we need to preprocess the data in order to standardize them.
python data_processsor.pyWe store common parameters combination inside exp.py. In order to run a the model with parameters combination exp1 inside exp.py. Simply run following command.
python main.py exp1The real meaning of each parameter is shown below.
'learning_rate_class' : the learning rate of classification task
'learning_rate_seg' : the learning rate of classification task
'training_iters' : the maximum iterations
'step_display' : the iterations before displaying the new accuracy
'step_save' : the iterations before saving the new model
'exp_name' : the name of experiment
'pretrainedStep' : the iterations of the pretrained model
'batch_size': batch size
'joint_ratio': the percentage of segmentation training
'joint_ratio_decay': whether use the join ratio deacy
'plot': whether plot the graph
'lr_decay': whether use the learning rate decay
'train' : whether train the model
'validation' : whether do validation
'test': whether do test
'selectedmodel': the name of model selected