instance-segmentation-with-discriminative-loss-tensorflow: A Python repository from qmind

Instance Segmentation with a Discriminative Loss Function

Tensorflow implementation of Semantic Instance Segmentation with a Discriminative Loss Function trained on the TuSimple dataset

Files

├── data here the data should be stored
│ └── tusimple_dataset_processing.py processes the TuSimple dataset
├── doc documentation
├── inference_test inference related data
│ └── images for testing the inference
├── trained_model pretrained model for finetuning
├── clustering.py mean-shift clustering
├── datagenerator.py feeds data for training and evaluation
├── enet.py Enet architecture
├── inference.py tests inference on images
├── loss.py defines discriminative loss function
├── README.md
├── training.py contains training pipeline
├── utils.py contains utilities files for building and initializing the graph
└── visualization.py contains visualization of the clustering and pixel embeddings

Instructions

Inference

To test the inference of the trained model execute:
python inference.py --modeldir trained_model --outdir inference_test/results

Training

Download the TuSimple training dataset and extract its contents to the data folder. The folder structure should look like this:
| data
├── train_set
│     ├── clips
│     ├── label_data_0313.json
│     ├── label_data_0531.json
│     ├── label_data_0601.json
│     └── readme.md
Run the following script to prepare images and labels.
python data/tusimple_dataset_processing.py <train_data_dir>
This should create the following images and labels folders:
| data
├── train_set
├── images
└── labels
For training on the dataset execute:
python training.py
alternatively use optional parameters (default parameters in this example):
python training --srcdir data --modeldir pretrained_semantic_model --outdir saved_model --logdir log --epochs 50 --var 1.0 --dist 1.0 --reg 1.0 --dvar 0.5 --ddist 1.5
To test the trained network execute: python inference.py --modeldir saved_model

Training Pipeline

Training Visualization

Feature space projection of one image for consecutive gradient steps. Each point represents one pixel embedding and each color represents an instance in the label.