1st Place Solution for Facebook's OpenEDS 2020 Sparse Semantic Segmentation Challenge

• Authors: Almaz Zinollayev, Zhanseri Ikram

• Video presentation for ECCV 2020: https://youtu.be/xXqDsPZyZic

• Competition: https://evalai.cloudcv.org/web/challenges/challenge-page/603/overview

Requirements

• Linux
• Python 3.6+
• PyTorch 1.3+
• CUDA 9.2+

To install requirements:

pip install -r requirements.txt

Data Preparation

First, prepare folds for training

python dataset/create_kfolds.py

Training

To train the model, run this command:

python train.py --config <path_to_config> --data-dir <path_to_data> --gpu <gpu_ids> --fold <fold_id>

List of other parameters can be seen inside train.py

Inference

To inference a model, run:

python predict_ensemble.py --data-dir <path_to_data> --label-file <path_to_file_with_sample_names> --save-dir <path_to_save> --gpu <gpu_ids>

Models that should be inferenced are written in predict_ensemble.py.
path_to_file_with_names is a file with all sample names e.g. S_0/label_0.npy and etc.

Evaluation

To evaluate model, run:

python validate.py --data-dir <path_to_data> --pred-dir <path_to_predictios> --fold <fold_id>

Predictions from model should be inferenced beforehand

Pre-trained Models

You can download pretrained models here:

• UNet EfficientNet-b7 trained on pseudolabels.

Results

Our models achieve the following performance on:

OpenEDS 2020 Sparse Semantic Segmentation Leaderboard

Model	mIoU
UNet EfficientNet-b7	95.15%
Ensemble of 10 Models	95.17%

The reason why the ensemble gives such a small accuracy increase is that most of the models in the ensemble were trained sub-optimally, as we've found out best training pipeline towards the end of the competition.

Acknowledgement

We thank Selim Seferbekov and Pavel Yakubovskiy for wonderful codebase that we extensively used during competition.