Semantic Instance Segmentation with a Discriminative Loss Function

This repository implements Semantic Instance Segmentation with a Discriminative Loss Function with some enhancements.

• Reference paper does not predict semantic segmentation mask, instead it uses ground-truth semantic segmentation mask. This code predicts semantic segmentation mask, similar to Towards End-to-End Lane Detection: an Instance Segmentation Approach.
• Reference paper predicts the number of instances implicity. It predicts embeddings for instances and predicts the number of instances as a result of clustering. Instead, this code predicts the number of instances as an output of network.
• Reference paper uses a segmentation network based on ResNet-38. Instead, this code uses either ReSeg with skip-connections based on first seven convolutional layers of VGG16 as segmentation network or an augmented version of Stacked Recurrent Hourglass.
• This code uses KMeans Clustering; however, reference paper uses "a fast variant of the mean-shift algorithm".

---

Modules

• Convolutional GRU
• Coordinate Convolution
  • AddCoordinates
  • CoordConv
  • CoordConvTranspose
  • CoordConvNet
• Recurrent Hourglass
• ReNet
• VGG16
  • VGG16
  • SkipVGG16

Networks

• ReSeg
• Stacked Recurrent Hourglass

---

In prediction phase, network inputs an image and outputs a semantic segmentation mask, the number of instances and embeddings for all pixels in the image. Then, foreground embeddings (which correspond to instances) are selected using semantic segmentation mask and foreground embeddings are clustered into "the number of instances" groups via clustering.

Installation

• Clone this repository : git clone --recursive https://github.com/Wizaron/instance-segmentation-pytorch.git
• Install ImageMagick : sudo apt install imagemagick
• Download and install Anaconda or Miniconda
• Create a conda environment : conda env create -f instance-segmentation-pytorch/code/conda_environment.yml

Data

CVPPP

• Download CVPPP dataset and extract downloaded zip file (CVPPP2017_LSC_training.zip) to instance-segmentation-pytorch/data/raw/CVPPP/
• This work uses A1 subset of the dataset.

Code Structure

• code: Codes for training and evaluation.
  • lib
    • lib/archs: Stores network architectures.
    • lib/archs/modules: Stores basic modules for architectures.
    • lib/model.py: Defines model (optimization, criterion, fit, predict, test, etc.).
    • lib/dataset.py: Data loading, augmentation, minibatching procedures.
    • lib/preprocess.py, lib/utils: Data augmentation methods.
    • lib/prediction.py: Prediction module.
    • lib/losses/dice.py: Dice loss for foreground semantic segmentation.
    • lib/losses/discriminative.py: Discriminative loss for instance segmentation.
  • settings
    • settings/CVPPP/data_settings.py: Defines settings about data.
    • settings/CVPPP/model_settings.py: Defines settings about model (hyper-parameters).
    • settings/CVPPP/training_settings.py: Defines settings for training (optimization method, weight decay, augmentation, etc.).
  • train.py: Training script.
  • pred.py: Prediction script for single image.
  • pred_list.py: Prediction scripts for a list of images.
  • evaluate.py: Evaluation script. Calculates SBD (symmetric best dice), |DiC| (absolute difference in count) and Foreground Dice (Dice score for semantic segmentation) as defined in the paper.
• data: Stores data and scripts to prepare dataset for training and evaluation.
  • metadata/CVPPP: Stores metadata; such as, training, validation and test splits, image shapes etc.
  • processed/CVPPP: Stores processed form of the data.
  • raw/CVPPP: Stores raw form of the data.
  • scripts: Stores scripts to prepare dataset.
    • scripts/CVPPP: For CVPPP dataset.
      • scripts/CVPPP/1-create_annotations.py: Saves annotations as a numpy array to processed/CVPPP/semantic-annotations/ and processed/CVPPP/instance-annotations.
      • scripts/CVPPP/1-remove_alpha.sh: Removes alpha channels from images. (In order to run this script, imagemagick should be installed.).
      • scripts/CVPPP/2-get_image_means-stds.py: Calculates and prints channel-wise means and standard deviations from training subset.
      • scripts/CVPPP/2-get_image_shapes.py: Saves image shapes to metadata/CVPPP/image_shapes.txt.
      • scripts/CVPPP/2-get_number_of_instances.py: Saves the number of instances in each image to metadata/CVPPP/number_of_instances.txt.
      • scripts/CVPPP/2-get_image_paths.py: Saves image paths to metadata/CVPPP/training_image_paths.txt, metadata/CVPPP/validation_image_paths.txt
      • scripts/CVPPP/3-create_dataset.py: Creates an lmdb dataset to processed/CVPPP/lmdb/. * scripts/CVPPP/prepare.sh: Runs the scripts above in a sequential manner.
• models/CVPPP: Stores checkpoints of the trained models.
• outputs/CVPPP: Stores predictions of the trained models.

Data Preparation

Data should be prepared prior to training and evaluation.

• Activate previously created conda environment : source activate ins-seg-pytorch or conda activate ins-seg-pytorch

CVPPP

• Place the extracted dataset to instance-segmentation-pytorch/data/raw/CVPPP/. Hence, raw dataset should be found at instance-segmentation-pytorch/data/raw/CVPPP/CVPPP2017_LSC_training/.
• In order to prepare the data go to instance-segmentation-pytorch/data/scripts/CVPPP/ and run sh prepare.sh.

Visdom Server

Start a Visdom server in a screen or tmux.

• Activate previously created conda environment : source activate ins-seg-pytorch or conda activate ins-seg-pytorch

• Start visdom server : python -m visdom.server

• We can access visdom server using http://localhost:8097

Training

• Activate previously created conda environment : source activate ins-seg-pytorch or conda activate ins-seg-pytorch

• Go to instance-segmentation-pytorch/code/ and run train.py.

usage: train.py [-h] [--model MODEL] [--usegpu] [--nepochs NEPOCHS]
                [--batchsize BATCHSIZE] [--debug] [--nworkers NWORKERS]
                --dataset DATASET

optional arguments:
  -h, --help            show this help message and exit
  --model MODEL         Filepath of trained model (to continue training)
                        [Default: '']
  --usegpu              Enables cuda to train on gpu [Default: False]
  --nepochs NEPOCHS     Number of epochs to train for [Default: 600]
  --batchsize BATCHSIZE
                        Batch size [Default: 2]
  --debug               Activates debug mode [Default: False]
  --nworkers NWORKERS   Number of workers for data loading (0 to do it using
                        main process) [Default : 2]
  --dataset DATASET     Name of the dataset which is "CVPPP"

Debug mode plots pixel embeddings to visdom, it reduces size of the embeddings to two-dimensions using TSNE. Hence, it slows training down.

As training continues, models are saved to instance-segmentation-pytorch/models/CVPPP.

Evaluation

After training is completed, we can make predictions.

• Activate previously created conda environment : source activate ins-seg-pytorch or conda activate ins-seg-pytorch

• Go to instance-segmentation-pytorch/code/.

• Run pred_list.py.

usage: pred_list.py [-h] --lst LST --model MODEL [--usegpu]
                    [--n_workers N_WORKERS] --dataset DATASET

optional arguments:
  -h, --help            show this help message and exit
  --lst LST             Text file that contains image paths
  --model MODEL         Path of the model
  --usegpu              Enables cuda to predict on gpu
  --dataset DATASET     Name of the dataset which is "CVPPP"

For example: python pred_list.py --lst ../data/metadata/CVPPP/validation_image_paths.txt --model ../models/CVPPP/2018-3-4_16-15_jcmaxwell_29-937494/model_155_0.123682662845.pth --usegpu --n_workers 4 --dataset CVPPP

• Predictions are written to outputs directory.
• After prediction is completed we can run evaluate.py. It prints output metrics to the stdout.

usage: evaluate.py [-h] --pred_dir PRED_DIR --dataset DATASET

optional arguments:
  -h, --help           show this help message and exit
  --pred_dir PRED_DIR  Prediction directory
  --dataset DATASET    Name of the dataset which is "CVPPP"

For example: python evaluate.py --pred_dir ../outputs/CVPPP/2018-3-4_16-15_jcmaxwell_29-937494-model_155_0.123682662845/validation/ --dataset CVPPP

Prediction

After training is complete, we can make predictions. We can use pred.py to make predictions for a single image.

• Activate previously created conda environment : source activate ins-seg-pytorch or conda activate ins-seg-pytorch

• Go to instance-segmentation-pytorch/code/.

• Run pred.py.

usage: pred.py [-h] --image IMAGE --model MODEL [--usegpu] --output OUTPUT
               [--n_workers N_WORKERS] --dataset DATASET

optional arguments:
  -h, --help            show this help message and exit
  --image IMAGE         Path of the image
  --model MODEL         Path of the model
  --usegpu              Enables cuda to predict on gpu
  --output OUTPUT       Path of the output directory
  --dataset DATASET     Name of the dataset which is "CVPPP"

Results

CVPPP

Scores on validation subset (28 images)

SBD	|DiC|	Foreground Dice
87.9	0.5	96.8

Sample Predictions

References

• VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION
• ReNet: A Recurrent Neural Network Based Alternative to Convolutional Networks
• DELVING DEEPER INTO CONVOLUTIONAL NETWORKS FOR LEARNING VIDEO REPRESENTATIONS
• ReSeg: A Recurrent Neural Network-based Model for Semantic Segmentation
• Semantic Instance Segmentation with a Discriminative Loss Function
• Instance Segmentation and Tracking with Cosine Embeddings and Recurrent Hourglass Networks
• An intriguing failing of convolutional neural networks and the CoordConv solution
• Leaf segmentation in plant phenotyping: A collation study