Need help with tensorflow-deeplab-lfov?
Click the “chat” button below for chat support from the developer who created it, or find similar developers for support.

About the developer

218 Stars 84 Forks MIT License 30 Commits 15 Opened issues


DeepLab-LargeFOV implemented in tensorflow

Services available


Need anything else?

Contributors list

# 19,958
30 commits


This is an implementation of DeepLab-LargeFOV in TensorFlow for semantic image segmentation on PASCAL VOC dataset.

Model Description

The DeepLab-LargeFOV is built on a fully convolutional variant of the VGG-16 net with several modifications: first, it exploits atrous (dilated) convolutions to increase the field-of-view; second, the number of filters in the last layers is reduced from 4096 to 1024 in order to decrease the memory consumption and the time spent on performing one forward-backward pass; third, it omits the last pooling layers to keep the downsampling ratio of 8.

The model is trained on a mini-batch of images and corresponding ground truth masks with the softmax classifier on the top. During training, the masks are downsampled to match the size of the output from the network; during inference, to acquire the output of the same size as the input, bilinear upsampling is applied. The final segmentation mask is acquired using argmax over unnormalised log scores from the network. Optionally, a fully-connected probabilistic graphical model, namely, CRF, can be applied to refine the final predictions. On the test set of PASCAL VOC, the model shows 70.3% of mean intersection-over-union.

For more details on the underlying model please refer to the following paper:

  title={DeepLab: Semantic Image Segmentation with Deep Convolutional Nets, Atrous Convolution, and Fully Connected CRFs},
  author={Liang-Chieh Chen and George Papandreou and Iasonas Kokkinos and Kevin Murphy and Alan L Yuille},


TensorFlow needs to be installed before running the scripts. TensorFlow>=0.11 is supported.

To install the required python packages (except TensorFlow), run

pip install -r requirements.txt
or for a local installation
pip install -user -r requirements.txt

Caffe to TensorFlow conversion

To imitate the structure of the model, we have used

files provided by the authors. The
script saves the structure of the network, i.e. the name of the parameters with their corresponding shapes (in TF 'HNWC' format), as well as the weights of those parameters (again, in the TF format). These weights can be used to initialise the variables in the model; otherwise, the filters will be initialised using the Xavier initialisation scheme, and biases will be initiliased as 0s. To use this script you will need to install Caffe. It is optional, and you can download two already converted models (
) here.


To train the network, we use the augmented PASCAL VOC 2012 dataset with 10582 images for training and 1449 images for validation.


We initialised the network from the

file provided by the authors. In that model, the last classification layer is randomly initialised using the Xavier scheme with biases set to zeros. The loss function is the pixel-wise softmax loss, and it is optimised using Adam. No weight decay is used.

script provides an ability to monitor model performance by snapshotting current results: Besides that, one can change the input size and augment data with random scaling.

To see the documentation on each of the training settings run the following:

python --help


After the training, the model shows 57% mIoU on the Pascal VOC 2012 validation dataset. The model initialised from the pre-trained

shows 67% mIoU on the same dataset. Note that in the original DeepLab each image is padded so that the input is of size 513x513 and CRF is used, which can be one of the reason of the lower score (~70.3% mIoU).

To see the documentation on each of the evaluation settings run the following:

python --help


To perform inference over your own images, use the following command:

python /path/to/your/image /path/to/ckpt/file
This will run the forward pass and save the resulted mask with this colour map:

Missing features

At the moment, the post-processing step with CRF is not implemented. Besides that, the weight decay is missing, as well.

Other implementations

We use cookies. If you continue to browse the site, you agree to the use of cookies. For more information on our use of cookies please see our Privacy Policy.