Semantic Segmentation Suite in TensorFlow. Implement, train, and test new Semantic Segmentation models easily!
Switch branches/tags
Nothing to show
Clone or download

Semantic Segmentation Suite in TensorFlow



What's New

  • Added the BiSeNet model from ECCV 2018!

  • Added the Dense Decoder Shortcut Connections model from CVPR 2018!

  • Added the DenseASPP model from CVPR 2018!

Coming Soon

Open up an issue to suggest a new feature or improvement!


This repository serves as a Semantic Segmentation Suite. The goal is to easily be able to implement, train, and test new Semantic Segmentation models! Complete with the following:

  • Training and testing modes
  • Data augmentation
  • Several state-of-the-art models. Easily plug and play with different models
  • Able to use any dataset
  • Evaluation including precision, recall, f1 score, average accuracy, per-class accuracy, and mean IoU
  • Plotting of loss function and accuracy over epochs

Any suggestions to improve this repository, including any new segmentation models you would like to see are welcome!

You can also check out my Transfer Learning Suite.


If you find this repository useful, please consider citing it using a link to the repo :)


The following feature extraction models are currently made available:


The following segmentation models are currently made available:

Files and Directories

  • Training on the dataset of your choice. Default is CamVid

  • Testing on the dataset of your choice. Default is CamVid

  • Use your newly trained model to run a prediction on a single image

  • Quick helper functions for data preparation and visualization

  • Utilities for printing, debugging, testing, and evaluation

  • models: Folder containing all model files. Use this to build your models, or use a pre-built one

  • CamVid: The CamVid datatset for Semantic Segmentation as a test bed. This is the 32 class version

  • checkpoints: Checkpoint files for each epoch during training

  • Test: Test results including images, per-class accuracies, precision, recall, and f1 score


This project has the following dependencies:

  • Numpy sudo pip install numpy

  • OpenCV Python sudo apt-get install python-opencv

  • TensorFlow sudo pip install --upgrade tensorflow-gpu


The only thing you have to do to get started is set up the folders in the following structure:

├── "dataset_name"                   
|   ├── train
|   ├── train_labels
|   ├── val
|   ├── val_labels
|   ├── test
|   ├── test_labels

Put a text file under the dataset directory called "class_dict.csv" which contains the list of classes along with the R, G, B colour labels to visualize the segmentation results. This kind of dictionairy is usually supplied with the dataset. Here is an example for the CamVid dataset:

Bicyclist,0,128, 192
Bridge,0, 128, 64
Building,128, 0, 0
Car,64, 0, 128
CartLuggagePram,64, 0, 192
Child,192, 128, 64
Column_Pole,192, 192, 128
Fence,64, 64, 128
LaneMkgsDriv,128, 0, 192
LaneMkgsNonDriv,192, 0, 64
Misc_Text,128, 128, 64
MotorcycleScooter,192, 0, 192
OtherMoving,128, 64, 64
ParkingBlock,64, 192, 128
Pedestrian,64, 64, 0
Road,128, 64, 128
RoadShoulder,128, 128, 192
Sidewalk,0, 0, 192
SignSymbol,192, 128, 128
Sky,128, 128, 128
SUVPickupTruck,64, 128,192
TrafficCone,0, 0, 64
TrafficLight,0, 64, 64
Train,192, 64, 128
Tree,128, 128, 0
Truck_Bus,192, 128, 192
Tunnel,64, 0, 64
VegetationMisc,192, 192, 0
Void,0, 0, 0
Wall,64, 192, 0

Note: If you are using any of the networks that rely on a pre-trained ResNet, then you will need to download the pre-trained weights using the provided script. These are currently: PSPNet, RefineNet, DeepLabV3, DeepLabV3+, GCN.

Then you can simply run! Check out the optional command line arguments:

usage: [-h] [--num_epochs NUM_EPOCHS]
                [--checkpoint_step CHECKPOINT_STEP]
                [--validation_step VALIDATION_STEP] [--image IMAGE]
                [--continue_training CONTINUE_TRAINING] [--dataset DATASET]
                [--crop_height CROP_HEIGHT] [--crop_width CROP_WIDTH]
                [--batch_size BATCH_SIZE] [--num_val_images NUM_VAL_IMAGES]
                [--h_flip H_FLIP] [--v_flip V_FLIP] [--brightness BRIGHTNESS]
                [--rotation ROTATION] [--model MODEL] [--frontend FRONTEND]

optional arguments:
  -h, --help            show this help message and exit
  --num_epochs NUM_EPOCHS
                        Number of epochs to train for
  --checkpoint_step CHECKPOINT_STEP
                        How often to save checkpoints (epochs)
  --validation_step VALIDATION_STEP
                        How often to perform validation (epochs)
  --image IMAGE         The image you want to predict on. Only valid in
                        "predict" mode.
  --continue_training CONTINUE_TRAINING
                        Whether to continue training from a checkpoint
  --dataset DATASET     Dataset you are using.
  --crop_height CROP_HEIGHT
                        Height of cropped input image to network
  --crop_width CROP_WIDTH
                        Width of cropped input image to network
  --batch_size BATCH_SIZE
                        Number of images in each batch
  --num_val_images NUM_VAL_IMAGES
                        The number of images to used for validations
  --h_flip H_FLIP       Whether to randomly flip the image horizontally for
                        data augmentation
  --v_flip V_FLIP       Whether to randomly flip the image vertically for data
  --brightness BRIGHTNESS
                        Whether to randomly change the image brightness for
                        data augmentation. Specifies the max bightness change
                        as a factor between 0.0 and 1.0. For example, 0.1
                        represents a max brightness change of 10% (+-).
  --rotation ROTATION   Whether to randomly rotate the image for data
                        augmentation. Specifies the max rotation angle in
  --model MODEL         The model you are using. See for
                        supported models
  --frontend FRONTEND   The frontend you are using. See
                        for supported models


These are some sample results for the CamVid dataset with 11 classes (previous research version).

In training, I used a batch size of 1 and image size of 352x480. The following results are for the FC-DenseNet103 model trained for 300 epochs. I used RMSProp with learning rate 0.001 and decay 0.995. I did not use any data augmentation like in the paper. I also didn't use any class balancing. These are just some quick and dirty example results.

Note that the checkpoint files are not uploaded to this repository since they are too big for GitHub (greater than 100 MB)

Class Original Accuracy My Accuracy
Sky 93.0 94.1
Building 83.0 81.2
Pole 37.8 38.3
Road 94.5 97.5
Pavement 82.2 87.9
Tree 77.3 75.5
SignSymbol 43.9 49.7
Fence 37.1 69.0
Car 77.3 87.0
Pedestrian 59.6 60.3
Bicyclist 50.5 75.3
Unlabelled N/A 40.9
Global 91.5 89.6
Loss vs Epochs Val. Acc. vs Epochs
alt text-1 alt text-2
Original GT Result
alt-text-3 alt-text-4 alt-text-5