AGarioud commited on
Commit
43f98e7
1 Parent(s): d051dc9

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +141 -1
README.md CHANGED
@@ -5,4 +5,144 @@ license_link: https://www.etalab.gouv.fr/wp-content/uploads/2018/11/open-licence
5
  pretty_name: French Land Cover from Aerospace Imagery
6
  size_categories:
7
  - 10B<n<100B
8
- ---
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
5
  pretty_name: French Land Cover from Aerospace Imagery
6
  size_categories:
7
  - 10B<n<100B
8
+ ---
9
+
10
+
11
+
12
+ # Challenge FLAIR #2: textural and temporal information for semantic segmentation from multi-source optical imagery
13
+
14
+ Code : [![license](https://img.shields.io/badge/License-Apache%202.0-blue.svg)](https://github.com/IGNF/FLAIR-1-AI-Challenge/blob/master/LICENSE) &emsp; Dataset : [![license](https://img.shields.io/badge/License-IO%202.0-green.svg)](https://github.com/etalab/licence-ouverte/blob/master/open-licence.md)
15
+
16
+ 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. Get in touch at ai-challenge@ign.fr
17
+
18
+
19
+ ![Alt bandeau FLAIR-IGN](images/flair_bandeau.jpg?raw=true)
20
+ <img width="100%" src="images/flair-2_logos-hd.png">
21
+
22
+ <div style="border-width:1px; border-style:solid; border-color:#d2db8c; padding-left: 1em; padding-right: 1em; ">
23
+
24
+ <h2 style="margin-top:5px;">Links</h2>
25
+
26
+
27
+ - **Datapaper : https://arxiv.org/pdf/2305.14467.pdf**
28
+
29
+ - **Dataset links :** https://ignf.github.io/FLAIR/#FLAIR2 [🛑 soon @ https://ignf.github.io/FLAIR/ !]
30
+
31
+ - **Challenge page : https://codalab.lisn.upsaclay.fr/competitions/13447**
32
+
33
+ </div>
34
+ <br><br>
35
+
36
+
37
+
38
+
39
+
40
+ ## Context & Data
41
+
42
+ The FLAIR #2 dataset is sampled countrywide and is composed of over 20 billion annotated pixels of very high resolution aerial imagery at 0.2 m spatial resolution, acquired over three years and different months (spatio-temporal domains). Aerial imagery patches consist of 5 channels (RVB-Near Infrared-Elevation) and have corresponding annotation (with 19 semantic classes or 13 for the baselines). Furthermore, to integrate broader spatial context and temporal information, high resolution Sentinel-2 1-year time series with 10 spectral band are also provided. More than 50,000 Sentinel-2 acquisitions with 10 m spatial resolution are available.
43
+ <br>
44
+
45
+ <p align="center">
46
+ <img width="40%" src="images/flair-2-spatial.png">
47
+ <br>
48
+ <em>Spatial definitions of the FLAIR #2 dataset.</em>
49
+ </p>
50
+
51
+
52
+ <p align="center">
53
+ <img width="85%" src="images/flair-2-patches.png">
54
+ <br>
55
+ <em>Example of input data (first three columns are from aerial imagery, fourth from Sentinel-2) and corresponding supervision masks (last column).</em>
56
+ </p>
57
+
58
+ <br><br>
59
+ ## Baseline model
60
+
61
+ A two-branch architecture integrating a U-Net <a href="https://github.com/qubvel/segmentation_models.pytorch"><img src="https://img.shields.io/badge/Link%20to-SMP-f4dbaa.svg"/></a> with a pre-trained ResNet34 encoder and a U-TAE <a href="https://github.com/VSainteuf/utae-paps"><img src="https://img.shields.io/badge/Link%20to-U--TAE-f4dbaa.svg"/></a> encompassing a temporal self-attention encoder is presented. The U-TAE branch aims at learning spatio-temporal embeddings from the high resolution satellite time series that are further integrated into the U-Net branch exploiting the aerial imagery. The proposed _U-T&T_ model features a fusion module to extend and reshape the U-TAE embeddings in order to add them towards the U-Net branch.
62
+
63
+ <p align="center">
64
+ <img width="100%" src="images/flair-2-network.png">
65
+ <br>
66
+ <em>Overview of the proposed two-branch architecture.</em>
67
+ </p>
68
+
69
+ <br><br>
70
+
71
+ ## Usage
72
+
73
+ The `flair-2-config.yml` file controls paths, hyperparameters and computing ressources. The file `requirement.txt` is listing used libraries for the baselines.
74
+
75
+ To launch a training/inference/metrics computation, you can either use :
76
+
77
+ - ```
78
+ main.py --config_file=flair-2-config.yml
79
+ ```
80
+
81
+ - use the `./notebook/flair-2-notebook.ipynb` notebook guiding you through data visualization, training and testing steps.
82
+
83
+ A toy dataset (reduced size) is available to check that your installation and the information in the configuration file are correct.
84
+
85
+ <br><br>
86
+
87
+ ## Leaderboard
88
+
89
+ Please note that for participants to the FLAIR #2 challenge on CodaLab, a certain number of constraints must be satisfied (in particular, inference time). All infos are available on the _Overview_ page of the competion.
90
+
91
+ | Model|Input|mIoU
92
+ ------------ | ------------- | -------------
93
+ | baseline U-Net (ResNet34) | aerial imagery | 0.5470
94
+ | baseline U-Net (ResNet34) + _metadata + augmentation_ | aerial imagery | 0.5593
95
+ |||
96
+ | baseline U-T&T | aerial and satellite imagery | 0.5594
97
+ | baseline U-T&T + _filter clouds + monthly averages + data augmentation_ | aerial and satellite imagery | 0.5758
98
+
99
+ If you want to submit a new entry, you can open a new issue.
100
+ <b> Results of the challenge will be reported after the end of the challenge early October! </b>
101
+
102
+ The baseline U-T&T + _filter clouds + monthly averages + data_augmentation_ obtains the following confusion matrix:
103
+
104
+ <br><br>
105
+ <p align="center">
106
+ <img width="50%" src="images/flair-2-confmat.png">
107
+ <br>
108
+ <em>Baseline confusion matrix of the test dataset normalized by rows.</em>
109
+ </p>
110
+
111
+
112
+ <br><br><br>
113
+
114
+ ## Reference
115
+ Please include a citation to the following article if you use the FLAIR #2 dataset:
116
+
117
+ ```
118
+ @article{ign2023flair2,
119
+ doi = {10.13140/RG.2.2.30938.93128/2},
120
+ url = {https://arxiv.org/pdf/2305.14467.pdf},
121
+ author = {Garioud, Anatol and {DE Wit}, Apolline and Poupée, Marc and Valette, Marion and Giordano, Sébastien and Wattrelos, Boris},
122
+ title = {FLAIR #2: textural and temporal information for semantic segmentation from multi-source optical imagery},
123
+ publisher = {arXiv},
124
+ year = {2023}
125
+ }
126
+ ```
127
+
128
+ ## Acknowledgment
129
+ This work was performed using HPC/AI resources from GENCI-IDRIS (Grant 2022-A0131013803). This work was supported by the project "Copernicus / FPCUP” of the European Union, by the French Space Agency (CNES) and by Connect by CNES.<br>
130
+
131
+
132
+
133
+
134
+
135
+ ## Dataset license
136
+
137
+ The "OPEN LICENCE 2.0/LICENCE OUVERTE" is a license created by the French government specifically for the purpose of facilitating the dissemination of open data by public administration.
138
+ If you are looking for an English version of this license, you can find it on the official GitHub page at the [official github page](https://github.com/etalab/licence-ouverte).
139
+
140
+ As stated by the license :
141
+
142
+ ### Applicable legislation
143
+
144
+ This licence is governed by French law.
145
+
146
+ ### Compatibility of this licence
147
+
148
+ This licence has been designed to be compatible with any free licence that at least requires an acknowledgement of authorship, and specifically with the previous version of this licence as well as with the following licences: United Kingdom’s “Open Government Licence” (OGL), Creative Commons’ “Creative Commons Attribution” (CC-BY) and Open Knowledge Foundation’s “Open Data Commons Attribution” (ODC-BY).