galoiszhang / weakly_localizations_tensorflow

CNNs based weakly supervised localizations in Tensorflow

Weakly Supervised Localization Implementations in Tensorflow

Tensorflow Implementations of

• Class Activation Maps
• Grad-CAM 'Gradient Based Class Activation Maps'
• Refining of results using Conditional Random Fields

Class Activation Maps

Class Activation Maps provides a way of doing weakly supervised localization in Convolutional Neural Networks. This tensorflow Implementation is partly inspired by this approach. Results obtained on Imagenet dataset is given below.

Note: The approach here uses the caffe weights provided by the author. The weights are converted from caffe to tensorflow's compatible weights using caffe-tensorflow.

• The model here uses VGG16 architecture.
• For Training you would need the pretrained weights.
• A link for the converted pretrained weight file will be provided soon.
• If you use the code then kindly cite the respective work

Grad-CAM (Gradient Based Class Activation Maps)

Grad-CAM is a recent state of the art approach for weakly supervised localization in CNNs. The general idea behind the approach is same as the CAM. However, instead of using linear combinations of last convolutional layer's feature maps with the output as weights to signify the importance of each feature map, they make use of the gradient backpropagations from the output to last convolutional layer feature maps. This strategy enables them to retain fully connected layer at the end of the network and hence it doesn't affect the classification accuracy.

• They also use guided backpropagation to get the pixelspace visualizations of the output prediction.
• The network architecture is VGG16.
• The weights are here are also converted from Caffe to Tensorflow using caffe-tensorflow.
• A link for the converted pretrained file weights will be provided soon.
• If you use the code then kindly cite the respective work

Refining of results using Conditional Random Fields

PyDenseCRF is a python wrapper of Phillip Krähenbühl's dense (fully connected) CRFs with gaussian edge potentials.

• For quick installation of the library

pip install git+https://github.com/lucasb-eyer/pydensecrf.git

• The provided cam_crf_notebook covers the basic usage of the library, however it is highly recommended to visit PyDenseCRF for detailed instructions.
• If you use the code then kindly cite the respective work

Efficient Inference in Fully Connected CRFs with Gaussian Edge Potentials
Philipp Krähenbühl and Vladlen Koltun
NIPS 2011

Requirements

• Python 3.5
• Tensorflow v0.12
• Opencv2