Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -18,13 +18,15 @@ tags:
|
|
18 |
|
19 |
## Context & Data
|
20 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
21 |
-
The hereby 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).
|
22 |
Aerial imagery patches consist of 5 channels (RVB-Near Infrared-Elevation) and have corresponding annotation (with 19 semantic classes or 13 for the baselines).
|
23 |
Furthermore, to integrate broader spatial context and temporal information, high resolution Sentinel-2 satellite 1-year time series with 10 spectral band are also provided.
|
24 |
More than 50,000 Sentinel-2 acquisitions with 10 m spatial resolution are available.
|
25 |
<br>
|
26 |
|
27 |
-
The dataset covers
|
|
|
|
|
28 |
|
29 |
<style type="text/css">
|
30 |
.tg {border-collapse:collapse;border-spacing:0;}
|
@@ -61,11 +63,13 @@ The dataset covers 50 spatial domains, encompassing 916 areas spanning 817 km².
|
|
61 |
<th class="tg-zv4m"></th>
|
62 |
<th class="tg-zv4m">Class</th>
|
63 |
<th class="tg-8jgo">Freq.-train (%)</th>
|
64 |
-
<th class="tg-8jgo">Freq.-test (%)</th>
|
|
|
65 |
<th class="tg-zv4m"></th>
|
66 |
<th class="tg-zv4m">Class</th>
|
67 |
<th class="tg-8jgo">Freq.-train (%)</th>
|
68 |
-
<th class="tg-8jgo">Freq.-test (%)</th>
|
|
|
69 |
</tr>
|
70 |
</thead>
|
71 |
<tbody>
|
@@ -73,96 +77,115 @@ The dataset covers 50 spatial domains, encompassing 916 areas spanning 817 km².
|
|
73 |
<td class="tg-2e1p"></td>
|
74 |
<td class="tg-km2t">(1) Building</td>
|
75 |
<td class="tg-8jgo">8.14</td>
|
|
|
76 |
<td class="tg-8jgo">3.26</td>
|
77 |
<td class="tg-l5fa"></td>
|
78 |
<td class="tg-km2t">(11) Agricultural Land</td>
|
79 |
<td class="tg-8jgo">10.98</td>
|
|
|
80 |
<td class="tg-8jgo">18.19</td>
|
81 |
</tr>
|
82 |
<tr>
|
83 |
<td class="tg-9efv"></td>
|
84 |
<td class="tg-km2t">(2) Pervious surface</td>
|
85 |
<td class="tg-8jgo">8.25</td>
|
|
|
86 |
<td class="tg-8jgo">3.82</td>
|
87 |
<td class="tg-rime"></td>
|
88 |
<td class="tg-km2t">(12) Plowed land</td>
|
89 |
<td class="tg-8jgo">3.88</td>
|
|
|
90 |
<td class="tg-8jgo">1.81</td>
|
91 |
</tr>
|
92 |
<tr>
|
93 |
<td class="tg-3m6m"></td>
|
94 |
<td class="tg-km2t">(3) Impervious surface</td>
|
95 |
<td class="tg-8jgo">13.72</td>
|
|
|
96 |
<td class="tg-8jgo">5.87</td>
|
97 |
<td class="tg-2cns"></td>
|
98 |
<td class="tg-km2t">(13) Swimming pool</td>
|
99 |
<td class="tg-8jgo">0.01</td>
|
|
|
100 |
<td class="tg-8jgo">0.02</td>
|
101 |
</tr>
|
102 |
<tr>
|
103 |
<td class="tg-r3rw"></td>
|
104 |
<td class="tg-km2t">(4) Bare soil</td>
|
105 |
<td class="tg-8jgo">3.47</td>
|
|
|
106 |
<td class="tg-8jgo">1.6</td>
|
107 |
<td class="tg-jjsp"></td>
|
108 |
<td class="tg-km2t">(14) Snow</td>
|
109 |
<td class="tg-8jgo">0.15</td>
|
|
|
110 |
<td class="tg-8jgo">-</td>
|
111 |
</tr>
|
112 |
<tr>
|
113 |
<td class="tg-9xgv"></td>
|
114 |
<td class="tg-km2t">(5) Water</td>
|
115 |
<td class="tg-8jgo">4.88</td>
|
|
|
116 |
<td class="tg-8jgo">3.17</td>
|
117 |
<td class="tg-2w6m"></td>
|
118 |
<td class="tg-km2t">(15) Clear cut</td>
|
119 |
<td class="tg-8jgo">0.15</td>
|
|
|
120 |
<td class="tg-8jgo">0.82</td>
|
121 |
</tr>
|
122 |
<tr>
|
123 |
<td class="tg-b45e"></td>
|
124 |
<td class="tg-km2t">(6) Coniferous</td>
|
125 |
<td class="tg-8jgo">2.74</td>
|
|
|
126 |
<td class="tg-8jgo">10.24</td>
|
127 |
<td class="tg-nla7"></td>
|
128 |
<td class="tg-km2t">(16) Mixed</td>
|
129 |
<td class="tg-8jgo">0.05</td>
|
|
|
130 |
<td class="tg-8jgo">0.12</td>
|
131 |
</tr>
|
132 |
<tr>
|
133 |
<td class="tg-qg2z"></td>
|
134 |
<td class="tg-km2t">(7) Deciduous</td>
|
135 |
<td class="tg-8jgo">15.38</td>
|
|
|
136 |
<td class="tg-8jgo">24.79</td>
|
137 |
<td class="tg-nv8o"></td>
|
138 |
<td class="tg-km2t">(17) Ligneous</td>
|
139 |
<td class="tg-8jgo">0.01</td>
|
|
|
140 |
<td class="tg-8jgo">-</td>
|
141 |
</tr>
|
142 |
<tr>
|
143 |
<td class="tg-grz5"></td>
|
144 |
<td class="tg-km2t">(8) Brushwood</td>
|
145 |
<td class="tg-8jgo">6.95</td>
|
146 |
-
<td class="tg-8jgo">
|
|
|
147 |
<td class="tg-bja1"></td>
|
148 |
<td class="tg-km2t">(18) Greenhouse</td>
|
149 |
<td class="tg-8jgo">0.12</td>
|
|
|
150 |
<td class="tg-8jgo">0.15</td>
|
151 |
</tr>
|
152 |
<tr>
|
153 |
<td class="tg-69kt"></td>
|
154 |
<td class="tg-km2t">(9) Vineyard</td>
|
155 |
<td class="tg-8jgo">3.13</td>
|
|
|
156 |
<td class="tg-8jgo">2.55</td>
|
157 |
<td class="tg-nto1"></td>
|
158 |
<td class="tg-km2t">(19) Other</td>
|
159 |
<td class="tg-8jgo">0.14</td>
|
|
|
160 |
<td class="tg-8jgo">0.04</td>
|
161 |
</tr>
|
162 |
<tr>
|
163 |
<td class="tg-r1r4"></td>
|
164 |
<td class="tg-km2t">(10) Herbaceous vegetation</td>
|
165 |
<td class="tg-8jgo">17.84</td>
|
|
|
166 |
<td class="tg-8jgo">19.76</td>
|
167 |
<td class="tg-zv4m"></td>
|
168 |
<td class="tg-zv4m"></td>
|
@@ -177,8 +200,10 @@ The dataset covers 50 spatial domains, encompassing 916 areas spanning 817 km².
|
|
177 |
|
178 |
## Dataset Structure
|
179 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
180 |
-
The FLAIR dataset consists of
|
181 |
-
|
|
|
|
|
182 |
|
183 |
<p align="center"><img src="readme_imgs/flair-patches.png" alt="" style="width:70%;max-width:600px;"/></p><br>
|
184 |
|
@@ -219,21 +244,23 @@ Satellite
|
|
219 |
Each pixel has been manually annotated by photo-interpretation of the 20 cm resolution aerial imagery, carried out by a team supervised by geography experts from the IGN.
|
220 |
Movable objects like cars or boats are annotated according to their underlying cover.
|
221 |
|
222 |
-
###
|
223 |
-
The dataset is made up of
|
224 |
-
For our experiments, we designate 32 domains for training, 8 for validation, and reserve 10
|
|
|
225 |
This arrangement ensures a balanced distribution of semantic classes, radiometric attributes, bioclimatic conditions, and acquisition times across each set.
|
226 |
Consequently, every split accurately reflects the landscape diversity inherent to metropolitan France.
|
227 |
-
It is important to mention that the patches come with meta-data permitting alternative splitting schemes
|
|
|
228 |
|
229 |
Official domain split: <br/>
|
230 |
|
231 |
<div style="display: flex; flex-wrap: nowrap; align-items: center">
|
232 |
<div style="flex: 40%;">
|
233 |
-
<img src="readme_imgs/flair-splits.png" alt="
|
234 |
</div>
|
235 |
|
236 |
-
<div style="flex: 60%;">
|
237 |
<table border="1">
|
238 |
<tr>
|
239 |
<th><font color="#c7254e">TRAIN:</font></th>
|
@@ -244,24 +271,53 @@ Official domain split: <br/>
|
|
244 |
<td>D004, D014, D029, D031, D058, D066, D067, D077</td>
|
245 |
</tr>
|
246 |
<tr>
|
247 |
-
<th><font color="#c7254e">TEST:</font></th>
|
248 |
-
<td>
|
249 |
</tr>
|
|
|
|
|
|
|
|
|
250 |
</table>
|
251 |
-
|
252 |
</div>
|
253 |
</div>
|
254 |
|
255 |
<br><br>
|
256 |
|
|
|
257 |
## Baseline code
|
258 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
259 |
We propose the U-T&T model, a two-branch architecture that combines spatial and temporal information from very high-resolution aerial images and high-resolution satellite images into a single output. The U-Net architecture is employed for the spatial/texture branch, using a ResNet34 backbone model pre-trained on ImageNet. For the spatio-temporal branch,
|
260 |
the U-TAE architecture incorporates a Temporal self-Attention Encoder (TAE) to explore the spatial and temporal characteristics of the Sentinel-2 time series data,
|
261 |
applying attention masks at different resolutions during decoding. This model allows for the fusion of learned information from both sources,
|
262 |
enhancing the representation of mono-date and time series data.
|
263 |
|
264 |
-
U-T&T code repository 📁 : https://github.com/IGNF/FLAIR-2
|
|
|
265 |
|
266 |
<th><font color="#c7254e"><b>IMPORTANT!</b></font></th> <b>The structure of the current dataset differs from the one that comes with the GitHub repository.</b>
|
267 |
To work with the current dataset, you need to replace the <font color=‘#D7881C’><em>src/load_data.py</em></font> file with the one provided here.
|
@@ -273,14 +329,8 @@ domains_train : ["D006_2020","D007_2020","D008_2019","D009_2019","D013_2020","D0
|
|
273 |
domains_val : ["D004_2021","D014_2020","D029_2021","D031_2019","D058_2020","D066_2021","D067_2021","D077_2021"]
|
274 |
domains_test : ["D015_2020","D022_2021","D026_2020","D036_2020","D061_2020","D064_2021","D068_2021","D069_2020","D071_2020","D084_2021"]
|
275 |
```
|
276 |
-
|
277 |
-
|
278 |
-
<br><br>
|
279 |
-
|
280 |
-
|
281 |
-
## Reference
|
282 |
-
<hr style='margin-top:-1em; margin-bottom:0' />
|
283 |
-
Please include a citation to the following article if you use the FLAIR dataset:
|
284 |
|
285 |
```
|
286 |
@inproceedings{garioud2023flair,
|
@@ -291,8 +341,28 @@ Please include a citation to the following article if you use the FLAIR dataset:
|
|
291 |
doi={https://doi.org/10.48550/arXiv.2310.13336},
|
292 |
}
|
293 |
```
|
|
|
294 |
|
295 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
296 |
|
297 |
## Acknowledgment
|
298 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
|
|
18 |
|
19 |
## Context & Data
|
20 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
21 |
+
The hereby FLAIR (#1 and #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).
|
22 |
Aerial imagery patches consist of 5 channels (RVB-Near Infrared-Elevation) and have corresponding annotation (with 19 semantic classes or 13 for the baselines).
|
23 |
Furthermore, to integrate broader spatial context and temporal information, high resolution Sentinel-2 satellite 1-year time series with 10 spectral band are also provided.
|
24 |
More than 50,000 Sentinel-2 acquisitions with 10 m spatial resolution are available.
|
25 |
<br>
|
26 |
|
27 |
+
The dataset covers 55 distinct spatial domains, encompassing 974 areas spanning 980 km². This dataset provides a robust foundation for advancing land cover mapping techniques.
|
28 |
+
We sample two test sets based on different input data and focus on semantic classes. The first test set (flair#1-test) uses very high resolution aerial imagery only and samples primarily anthropized land cover classes.
|
29 |
+
In contrast, the second test set (flair#2-test) combines aerial and satellite imagery and has more natural classes with temporal variations represented.<br><br>
|
30 |
|
31 |
<style type="text/css">
|
32 |
.tg {border-collapse:collapse;border-spacing:0;}
|
|
|
63 |
<th class="tg-zv4m"></th>
|
64 |
<th class="tg-zv4m">Class</th>
|
65 |
<th class="tg-8jgo">Freq.-train (%)</th>
|
66 |
+
<th class="tg-8jgo">Freq.-test flair#1 (%)</th>
|
67 |
+
<th class="tg-8jgo">Freq.-test flair#2 (%)</th>
|
68 |
<th class="tg-zv4m"></th>
|
69 |
<th class="tg-zv4m">Class</th>
|
70 |
<th class="tg-8jgo">Freq.-train (%)</th>
|
71 |
+
<th class="tg-8jgo">Freq.-test flair#1 (%)</th>
|
72 |
+
<th class="tg-8jgo">Freq.-test flair#2 (%)</th>
|
73 |
</tr>
|
74 |
</thead>
|
75 |
<tbody>
|
|
|
77 |
<td class="tg-2e1p"></td>
|
78 |
<td class="tg-km2t">(1) Building</td>
|
79 |
<td class="tg-8jgo">8.14</td>
|
80 |
+
<td class="tg-8jgo">8.6</td>
|
81 |
<td class="tg-8jgo">3.26</td>
|
82 |
<td class="tg-l5fa"></td>
|
83 |
<td class="tg-km2t">(11) Agricultural Land</td>
|
84 |
<td class="tg-8jgo">10.98</td>
|
85 |
+
<td class="tg-8jgo">6.95</td>
|
86 |
<td class="tg-8jgo">18.19</td>
|
87 |
</tr>
|
88 |
<tr>
|
89 |
<td class="tg-9efv"></td>
|
90 |
<td class="tg-km2t">(2) Pervious surface</td>
|
91 |
<td class="tg-8jgo">8.25</td>
|
92 |
+
<td class="tg-8jgo">7.34</td>
|
93 |
<td class="tg-8jgo">3.82</td>
|
94 |
<td class="tg-rime"></td>
|
95 |
<td class="tg-km2t">(12) Plowed land</td>
|
96 |
<td class="tg-8jgo">3.88</td>
|
97 |
+
<td class="tg-8jgo">2.25</td>
|
98 |
<td class="tg-8jgo">1.81</td>
|
99 |
</tr>
|
100 |
<tr>
|
101 |
<td class="tg-3m6m"></td>
|
102 |
<td class="tg-km2t">(3) Impervious surface</td>
|
103 |
<td class="tg-8jgo">13.72</td>
|
104 |
+
<td class="tg-8jgo">14.98</td>
|
105 |
<td class="tg-8jgo">5.87</td>
|
106 |
<td class="tg-2cns"></td>
|
107 |
<td class="tg-km2t">(13) Swimming pool</td>
|
108 |
<td class="tg-8jgo">0.01</td>
|
109 |
+
<td class="tg-8jgo">0.04</td>
|
110 |
<td class="tg-8jgo">0.02</td>
|
111 |
</tr>
|
112 |
<tr>
|
113 |
<td class="tg-r3rw"></td>
|
114 |
<td class="tg-km2t">(4) Bare soil</td>
|
115 |
<td class="tg-8jgo">3.47</td>
|
116 |
+
<td class="tg-8jgo">4.36</td>
|
117 |
<td class="tg-8jgo">1.6</td>
|
118 |
<td class="tg-jjsp"></td>
|
119 |
<td class="tg-km2t">(14) Snow</td>
|
120 |
<td class="tg-8jgo">0.15</td>
|
121 |
+
<td class="tg-8jgo">-</td>
|
122 |
<td class="tg-8jgo">-</td>
|
123 |
</tr>
|
124 |
<tr>
|
125 |
<td class="tg-9xgv"></td>
|
126 |
<td class="tg-km2t">(5) Water</td>
|
127 |
<td class="tg-8jgo">4.88</td>
|
128 |
+
<td class="tg-8jgo">5.98</td>
|
129 |
<td class="tg-8jgo">3.17</td>
|
130 |
<td class="tg-2w6m"></td>
|
131 |
<td class="tg-km2t">(15) Clear cut</td>
|
132 |
<td class="tg-8jgo">0.15</td>
|
133 |
+
<td class="tg-8jgo">0.01</td>
|
134 |
<td class="tg-8jgo">0.82</td>
|
135 |
</tr>
|
136 |
<tr>
|
137 |
<td class="tg-b45e"></td>
|
138 |
<td class="tg-km2t">(6) Coniferous</td>
|
139 |
<td class="tg-8jgo">2.74</td>
|
140 |
+
<td class="tg-8jgo">2.39</td>
|
141 |
<td class="tg-8jgo">10.24</td>
|
142 |
<td class="tg-nla7"></td>
|
143 |
<td class="tg-km2t">(16) Mixed</td>
|
144 |
<td class="tg-8jgo">0.05</td>
|
145 |
+
<td class="tg-8jgo">-</td>
|
146 |
<td class="tg-8jgo">0.12</td>
|
147 |
</tr>
|
148 |
<tr>
|
149 |
<td class="tg-qg2z"></td>
|
150 |
<td class="tg-km2t">(7) Deciduous</td>
|
151 |
<td class="tg-8jgo">15.38</td>
|
152 |
+
<td class="tg-8jgo">13.91</td>
|
153 |
<td class="tg-8jgo">24.79</td>
|
154 |
<td class="tg-nv8o"></td>
|
155 |
<td class="tg-km2t">(17) Ligneous</td>
|
156 |
<td class="tg-8jgo">0.01</td>
|
157 |
+
<td class="tg-8jgo">0.03</td>
|
158 |
<td class="tg-8jgo">-</td>
|
159 |
</tr>
|
160 |
<tr>
|
161 |
<td class="tg-grz5"></td>
|
162 |
<td class="tg-km2t">(8) Brushwood</td>
|
163 |
<td class="tg-8jgo">6.95</td>
|
164 |
+
<td class="tg-8jgo">6.91</td>
|
165 |
+
<td class="tg-8jgo">3.81</td>
|
166 |
<td class="tg-bja1"></td>
|
167 |
<td class="tg-km2t">(18) Greenhouse</td>
|
168 |
<td class="tg-8jgo">0.12</td>
|
169 |
+
<td class="tg-8jgo">0.2</td>
|
170 |
<td class="tg-8jgo">0.15</td>
|
171 |
</tr>
|
172 |
<tr>
|
173 |
<td class="tg-69kt"></td>
|
174 |
<td class="tg-km2t">(9) Vineyard</td>
|
175 |
<td class="tg-8jgo">3.13</td>
|
176 |
+
<td class="tg-8jgo">3.87</td>
|
177 |
<td class="tg-8jgo">2.55</td>
|
178 |
<td class="tg-nto1"></td>
|
179 |
<td class="tg-km2t">(19) Other</td>
|
180 |
<td class="tg-8jgo">0.14</td>
|
181 |
+
<td class="tg-8jgo">0.-</td>
|
182 |
<td class="tg-8jgo">0.04</td>
|
183 |
</tr>
|
184 |
<tr>
|
185 |
<td class="tg-r1r4"></td>
|
186 |
<td class="tg-km2t">(10) Herbaceous vegetation</td>
|
187 |
<td class="tg-8jgo">17.84</td>
|
188 |
+
<td class="tg-8jgo">22.17</td>
|
189 |
<td class="tg-8jgo">19.76</td>
|
190 |
<td class="tg-zv4m"></td>
|
191 |
<td class="tg-zv4m"></td>
|
|
|
200 |
|
201 |
## Dataset Structure
|
202 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
203 |
+
The FLAIR dataset consists of a total of 93 462 patches: 61 712 patches for the train/val dataset, 15 700 patches for flair#1-test and 16 050 patches for flair#2-test.
|
204 |
+
|
205 |
+
Each patch includes a high-resolution aerial image (512x512) at 0.2 m, a yearly satellite image time series (40x40 by default by wider areas are provided) with a spatial resolution of 10 m
|
206 |
+
and associated cloud and snow masks (available in train/val and flair#2-test), and pixel-precise elevation and land cover annotations at 0.2 m resolution (512x512).
|
207 |
|
208 |
<p align="center"><img src="readme_imgs/flair-patches.png" alt="" style="width:70%;max-width:600px;"/></p><br>
|
209 |
|
|
|
244 |
Each pixel has been manually annotated by photo-interpretation of the 20 cm resolution aerial imagery, carried out by a team supervised by geography experts from the IGN.
|
245 |
Movable objects like cars or boats are annotated according to their underlying cover.
|
246 |
|
247 |
+
### Data Splits
|
248 |
+
The dataset is made up of 55 distinct spatial domains, aligned with the administrative boundaries of the French départements.
|
249 |
+
For our experiments, we designate 32 domains for training, 8 for validation, and reserve 10 official test sets for flair#1-test and flair#2-test.
|
250 |
+
It can also be noted that some domains are common in the flair#1-test and flair#2-test datasets but cover different areas within the domain.
|
251 |
This arrangement ensures a balanced distribution of semantic classes, radiometric attributes, bioclimatic conditions, and acquisition times across each set.
|
252 |
Consequently, every split accurately reflects the landscape diversity inherent to metropolitan France.
|
253 |
+
It is important to mention that the patches come with meta-data permitting alternative splitting schemes.
|
254 |
+
|
255 |
|
256 |
Official domain split: <br/>
|
257 |
|
258 |
<div style="display: flex; flex-wrap: nowrap; align-items: center">
|
259 |
<div style="flex: 40%;">
|
260 |
+
<img src="readme_imgs/flair-splits.png" alt="flair-splits">
|
261 |
</div>
|
262 |
|
263 |
+
<div style="flex: 60%; margin: auto;"">
|
264 |
<table border="1">
|
265 |
<tr>
|
266 |
<th><font color="#c7254e">TRAIN:</font></th>
|
|
|
271 |
<td>D004, D014, D029, D031, D058, D066, D067, D077</td>
|
272 |
</tr>
|
273 |
<tr>
|
274 |
+
<th><font color="#c7254e">TEST-flair#1:</font></th>
|
275 |
+
<td>D012, D022, D026, D064, D068, D071, D075, D076, D083, D085</td>
|
276 |
</tr>
|
277 |
+
<tr>
|
278 |
+
<th><font color="#c7254e">TEST-flair#2:</font></th>
|
279 |
+
<td>D015, D022, D026, D036, D061, D064, D068, D069, D071, D084</td>
|
280 |
+
</tr>
|
281 |
</table>
|
|
|
282 |
</div>
|
283 |
</div>
|
284 |
|
285 |
<br><br>
|
286 |
|
287 |
+
|
288 |
## Baseline code
|
289 |
<hr style='margin-top:-1em; margin-bottom:0' />
|
290 |
+
<br>
|
291 |
+
|
292 |
+
### Flair #1 (aerial only)
|
293 |
+
A U-Net architecture with a pre-trained ResNet34 encoder from the pytorch segmentation models library is used for the baselines.
|
294 |
+
The used architecture allows integration of patch-wise metadata information and employs commonly used image data augmentation techniques.
|
295 |
+
|
296 |
+
Flair#1 code repository 📁 : https://github.com/IGNF/FLAIR-1<br/>
|
297 |
+
Link to the paper : https://arxiv.org/pdf/2211.12979.pdf <br>
|
298 |
+
|
299 |
+
Please include a citation to the following article if you use the FLAIR#1 dataset:
|
300 |
+
|
301 |
+
```
|
302 |
+
@article{ign2022flair1,
|
303 |
+
doi = {10.13140/RG.2.2.30183.73128/1},
|
304 |
+
url = {https://arxiv.org/pdf/2211.12979.pdf},
|
305 |
+
author = {Garioud, Anatol and Peillet, Stéphane and Bookjans, Eva and Giordano, Sébastien and Wattrelos, Boris},
|
306 |
+
title = {FLAIR #1: semantic segmentation and domain adaptation dataset},
|
307 |
+
publisher = {arXiv},
|
308 |
+
year = {2022}
|
309 |
+
}
|
310 |
+
```
|
311 |
+
<br>
|
312 |
+
|
313 |
+
### Flair #2 (aerial and satellite)
|
314 |
We propose the U-T&T model, a two-branch architecture that combines spatial and temporal information from very high-resolution aerial images and high-resolution satellite images into a single output. The U-Net architecture is employed for the spatial/texture branch, using a ResNet34 backbone model pre-trained on ImageNet. For the spatio-temporal branch,
|
315 |
the U-TAE architecture incorporates a Temporal self-Attention Encoder (TAE) to explore the spatial and temporal characteristics of the Sentinel-2 time series data,
|
316 |
applying attention masks at different resolutions during decoding. This model allows for the fusion of learned information from both sources,
|
317 |
enhancing the representation of mono-date and time series data.
|
318 |
|
319 |
+
U-T&T code repository 📁 : https://github.com/IGNF/FLAIR-2<br/>
|
320 |
+
Link to the paper : https://arxiv.org/abs/2310.13336 <br>
|
321 |
|
322 |
<th><font color="#c7254e"><b>IMPORTANT!</b></font></th> <b>The structure of the current dataset differs from the one that comes with the GitHub repository.</b>
|
323 |
To work with the current dataset, you need to replace the <font color=‘#D7881C’><em>src/load_data.py</em></font> file with the one provided here.
|
|
|
329 |
domains_val : ["D004_2021","D014_2020","D029_2021","D031_2019","D058_2020","D066_2021","D067_2021","D077_2021"]
|
330 |
domains_test : ["D015_2020","D022_2021","D026_2020","D036_2020","D061_2020","D064_2021","D068_2021","D069_2020","D071_2020","D084_2021"]
|
331 |
```
|
332 |
+
<br>
|
333 |
+
Please include a citation to the following article if you use the FLAIR#2 dataset:
|
|
|
|
|
|
|
|
|
|
|
|
|
334 |
|
335 |
```
|
336 |
@inproceedings{garioud2023flair,
|
|
|
341 |
doi={https://doi.org/10.48550/arXiv.2310.13336},
|
342 |
}
|
343 |
```
|
344 |
+
<br>
|
345 |
|
346 |
+
## CodaLab challenges
|
347 |
+
<hr style='margin-top:-1em; margin-bottom:0' />
|
348 |
+
|
349 |
+
The FLAIR dataset was used for two challenges organized by IGN in 2023 on the CodaLab platform.<br>
|
350 |
+
Challenge FLAIR#1 : https://codalab.lisn.upsaclay.fr/competitions/8769 <br>
|
351 |
+
Challenge FLAIR#2 : https://codalab.lisn.upsaclay.fr/competitions/13447 <br>
|
352 |
+
<br>
|
353 |
+
|
354 |
+
flair#1-test | The podium:
|
355 |
+
🥇 businiao - 0.65920
|
356 |
+
🥈 Breizhchess - 0.65600
|
357 |
+
🥉 wangzhiyu918 - 0.64930
|
358 |
+
|
359 |
+
flair#2-test | The podium:
|
360 |
+
The podium:
|
361 |
+
🥇 strakajk - 0.64130
|
362 |
+
🥈 Breizhchess - 0.63550
|
363 |
+
🥉 qwerty64 - 0.63510
|
364 |
+
|
365 |
+
<br>
|
366 |
|
367 |
## Acknowledgment
|
368 |
<hr style='margin-top:-1em; margin-bottom:0' />
|