VGG16-places365 model for scene classification, written in Keras 2.0

You have just found the Keras (2.0) implementation of the pre-trained VGG 16 convolutional-layer CNN model on Places365-Standard (~1.8 million images from 365 scene categories).

Overview

CNNs trained on Places365 database (latest subset of Places2 Database) could be directly used for scene recognition, while the deep scene features from the higher level layer of CNN could be used as generic features for visual recognition.

Paper

The Keras models has been obtained by directly converting the Caffe model provived by the authors (all the original Caffe-based resources can be found there).

More details about the network architecture can be found in the following paper:

Places: A 10 million Image Database for Scene Recognition
Zhou, B., Lapedriza, A., Khosla, A., Oliva, A., & Torralba, A.
IEEE Transactions on Pattern Analysis and Machine Intelligence

Please consider citing the paper if you use the pre-trained CNN models.

Contents:

Models:

• vgg16_places_365.py
• vgg16_hybrid_places_1365.py

Weights:

• VGG16_Places365_weights.h5
• VGG16_Hybrid_Places1365_weights.h5

Serialised JSONs:

• VGG16_Places365.json
• VGG16_Hybrid_Places1365.json

Usage: Download the weights which were dumped out from the initial Caffe resources and set the argument weights_path accordingly in vgg16_places_365.py & vgg16_hybrid_places_1365.py

Keras Model Summary:

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
data (InputLayer)            (None, 3, 224, 224)       0         
_________________________________________________________________
conv1_1 (Conv2D)             (None, 64, 224, 224)      1792      
_________________________________________________________________
conv1_2 (Conv2D)             (None, 64, 224, 224)      36928     
_________________________________________________________________
pool1 (MaxPooling2D)         (None, 64, 112, 112)      0         
_________________________________________________________________
conv2_1 (Conv2D)             (None, 128, 112, 112)     73856     
_________________________________________________________________
conv2_2 (Conv2D)             (None, 128, 112, 112)     147584    
_________________________________________________________________
pool2 (MaxPooling2D)         (None, 128, 56, 56)       0         
_________________________________________________________________
conv3_1 (Conv2D)             (None, 256, 56, 56)       295168    
_________________________________________________________________
conv3_2 (Conv2D)             (None, 256, 56, 56)       590080    
_________________________________________________________________
conv3_3 (Conv2D)             (None, 256, 56, 56)       590080    
_________________________________________________________________
pool3 (MaxPooling2D)         (None, 256, 28, 28)       0         
_________________________________________________________________
conv4_1 (Conv2D)             (None, 512, 28, 28)       1180160   
_________________________________________________________________
conv4_2 (Conv2D)             (None, 512, 28, 28)       2359808   
_________________________________________________________________
conv4_3 (Conv2D)             (None, 512, 28, 28)       2359808   
_________________________________________________________________
pool4 (MaxPooling2D)         (None, 512, 14, 14)       0         
_________________________________________________________________
conv5_1 (Conv2D)             (None, 512, 14, 14)       2359808   
_________________________________________________________________
conv5_2 (Conv2D)             (None, 512, 14, 14)       2359808   
_________________________________________________________________
conv5_3 (Conv2D)             (None, 512, 14, 14)       2359808   
_________________________________________________________________
pool5 (MaxPooling2D)         (None, 512, 7, 7)         0         
_________________________________________________________________
flatten (Flatten)            (None, 25088)             0         
_________________________________________________________________
fc6 (Dense)                  (None, 4096)              102764544 
_________________________________________________________________
drop6 (Dropout)              (None, 4096)              0         
_________________________________________________________________
fc7 (Dense)                  (None, 4096)              16781312  
_________________________________________________________________
drop7 (Dropout)              (None, 4096)              0         
_________________________________________________________________
fc8a (Dense)                 (None, 365)               1495405   
_________________________________________________________________
prob (Activation)            (None, 365)               0         
=================================================================
Total params: 135,755,949
Trainable params: 135,755,949
Non-trainable params: 0

Licensing

We are always interested in how this model is being used, so if you found this model useful or plan to make a release of code based on or using this package, it would be great to hear from you.

Other Models

This is going to be an evolving repository and I will keep updating it with Keras-compatible models which are not included in Keras Applications, so make sure you have starred and forked this repository before moving on !

Questions and Comments

If you have any suggestions or bugs to report you can pull a request or start a discussion.

---