【数据科学赛】FLAIR #1: Semantic segmentation and domain adaptation #CV #$10,000 #CodaLab

CompHub 实时聚合多平台的数据类(Kaggle、天池…)和OJ类(Leetcode、牛客…）比赛。本账号会推送最新的比赛消息，欢迎关注！

更多比赛信息见 CompHub主页 或 点击文末阅读原文

---

以下内容摘录自比赛主页

Part1赛题介绍

题目

FLAIR #1: Semantic segmentation and domain adaptation

主办方

French National Institute of Geographical and Forest Information (IGN).

举办平台

CodaLab

背景

Participate in obtaining more accurate maps for a more comprehensive description and a better understanding of our environment! Come push the limits of state-of-the-art semantic segmentation approaches on a large and challenging dataset.

Part2时间安排

• Start: Nov. 21, 2022, 10 a.m.

• End: March 21, 2023, 11:59 p.m.

Part3奖励机制

The award winners will be announced once the competition ends beginning of April. The distribution of the prizes is as follows:

• 🥇 1st place = 5000 EUR

• 🥈 2nd place = 3500 EUR

• 🥉 3rd place = 1500 EUR

Part4赛题描述

We present here a large dataset ( >20 billion pixels) of aerial imagery, topographic information and land cover (buildings, water, forest, agriculture...) annotations with the aim to further advance research on semantic segmentation , domain adaptation and transfer learning. Countrywide remote sensing aerial imagery is by necessity acquired at different times and dates and under different conditions. Likewise, at large scales, the characteristics of semantic classes can vary depending on location and become heterogenous. This opens up challenges for the spatial and temporal generalization of deep learning models!

ORTHO HR® aerial image cover of France (left) and train and test spatial domains of the dataset (right).

Part5比赛数据

The FLAIR-one dataset consists of 77,412 high resolution (0.2 m spatial resolution) patches with 13 semantic classes (19 original classes remapped to 13, see the associated paper in the starting kit for explanation). The dataset covers a total of approximatly 800 km², with patches that have been sampled accross the entire metropolitan French territory to be illustrating the different climate and landscapes (spatial domains). The aerial images included in the dataset were acquired during different months and years (temporal domains).

A U-Net architecture with a pre-trained ResNet34 encoder from the pytorch segmentation models library is used for the baselines. The used architecture allows integration of patch-wise metadata information and employs commonly used image data augmentation techniques. Results are presented in the technical description of the dataset.

Example of input data (first three columns) and corresponding supervision masks (last column).

Part6评测指标

Evaluation of the participant submission (inference of their model on the test set containing 15,700 patches) is made through the mean Intersection-over-Union (mIoU).

---