Update README.md
Browse files
README.md
CHANGED
@@ -79,7 +79,7 @@ pipeline_tag: image-segmentation
|
|
79 |
|
80 |
# FLAIR model collection
|
81 |
The FLAIR models is a collection of semantic segmentation models initially developed to classify land cover on very high resolution aerial ortho-images ([BD ORTHO®](https://geoservices.ign.fr/bdortho)).
|
82 |
-
The distributed pre-trained
|
83 |
- input modalities : RVB (true colours), RVBI (true colours + infrared), RVBIE (true colours + infrared + elevation)
|
84 |
- model architecture : U-Net with a Resnet-34 encoder, Deeplab
|
85 |
- target class nomenclature : 12 or 15 land cover classes
|
@@ -93,7 +93,6 @@ The general characteristics of this specific model *FLAIR-INC_RVBIE_resnet34_une
|
|
93 |
* 15 class nomenclature [building,pervious_surface,impervious_surface,bare_soil,water,coniferous,deciduous,brushwood,vineyard,herbaceous,agricultural_land,plowed_land,swimming pool,snow,greenhouse]
|
94 |
|
95 |
|
96 |
-
|
97 |
## Model Informations
|
98 |
|
99 |
<!-- Provide the basic links for the model. -->
|
@@ -106,25 +105,24 @@ The general characteristics of this specific model *FLAIR-INC_RVBIE_resnet34_une
|
|
106 |
- hardware: GENCI, XXX
|
107 |
- **License:** : Apache 2.0
|
108 |
|
109 |
-
|
110 |
|
111 |
## Uses
|
112 |
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
113 |
|
114 |
The model has been trained with
|
115 |
|
116 |
-
|
117 |
**Radiometry of input images** :
|
118 |
The input images are distributed in 8-bit encoding format per channel. or traning the model, input normalization was performed (see section **Traing Details**).
|
119 |
It is recommended that the user apply the same type of input normalization while inferring the model.
|
120 |
|
121 |
-
|
122 |
**Multi-domain model** :
|
123 |
The FLAIR-INC dataset that was used for training is composed of 75 radiometric domains. In the case of aerial images, domain shifts are due : the date of the aerial survey (april to november), spatial domain (equivalent to a french department administrative division) and downstream radimetric processing.
|
124 |
By construction the model is robust to theses shifts, and can be applied to any images of the ([BD ORTHO® product](https://geoservices.ign.fr/bdortho)).
|
125 |
|
126 |
**Land Cover classes of prediction** :
|
127 |
-
The orginial class nomenclature of the FLAIR Dataset is made up of 19 classes
|
|
|
|
|
128 |
|
129 |
|
130 |
## Bias, Risks, and Limitations
|
@@ -144,8 +142,6 @@ The FLAIR-INC_RVBIE_resnet34_unet_15cl_norm model has been trained on patches re
|
|
144 |
The user should be aware that applying the model to other type of landscapes may imply a drop in model metrics.
|
145 |
|
146 |
|
147 |
-
|
148 |
-
|
149 |
{{ bias_risks_limitations | default("[More Information Needed]", true)}}
|
150 |
|
151 |
|
@@ -160,6 +156,7 @@ The user should be aware that applying the model to other type of landscapes may
|
|
160 |
|
161 |
Use the code below to get started with the model.
|
162 |
|
|
|
163 |
{{ get_started_code | default("[More Information Needed]", true)}}
|
164 |
|
165 |
## Training Details
|
@@ -199,6 +196,38 @@ Input normalization was performed
|
|
199 |
|
200 |
- **Training regime:** {{ training_regime | default("[More Information Needed]", true)}} <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
|
201 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
202 |
#### Speeds, Sizes, Times [optional]
|
203 |
|
204 |
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|
|
|
79 |
|
80 |
# FLAIR model collection
|
81 |
The FLAIR models is a collection of semantic segmentation models initially developed to classify land cover on very high resolution aerial ortho-images ([BD ORTHO®](https://geoservices.ign.fr/bdortho)).
|
82 |
+
The distributed pre-trained models differ in their :
|
83 |
- input modalities : RVB (true colours), RVBI (true colours + infrared), RVBIE (true colours + infrared + elevation)
|
84 |
- model architecture : U-Net with a Resnet-34 encoder, Deeplab
|
85 |
- target class nomenclature : 12 or 15 land cover classes
|
|
|
93 |
* 15 class nomenclature [building,pervious_surface,impervious_surface,bare_soil,water,coniferous,deciduous,brushwood,vineyard,herbaceous,agricultural_land,plowed_land,swimming pool,snow,greenhouse]
|
94 |
|
95 |
|
|
|
96 |
## Model Informations
|
97 |
|
98 |
<!-- Provide the basic links for the model. -->
|
|
|
105 |
- hardware: GENCI, XXX
|
106 |
- **License:** : Apache 2.0
|
107 |
|
|
|
108 |
|
109 |
## Uses
|
110 |
<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
|
111 |
|
112 |
The model has been trained with
|
113 |
|
|
|
114 |
**Radiometry of input images** :
|
115 |
The input images are distributed in 8-bit encoding format per channel. or traning the model, input normalization was performed (see section **Traing Details**).
|
116 |
It is recommended that the user apply the same type of input normalization while inferring the model.
|
117 |
|
|
|
118 |
**Multi-domain model** :
|
119 |
The FLAIR-INC dataset that was used for training is composed of 75 radiometric domains. In the case of aerial images, domain shifts are due : the date of the aerial survey (april to november), spatial domain (equivalent to a french department administrative division) and downstream radimetric processing.
|
120 |
By construction the model is robust to theses shifts, and can be applied to any images of the ([BD ORTHO® product](https://geoservices.ign.fr/bdortho)).
|
121 |
|
122 |
**Land Cover classes of prediction** :
|
123 |
+
The orginial class nomenclature of the FLAIR Dataset is made up of 19 classes(See the [FLAIR dataset](https://huggingface.co/datasets/IGNF/FLAIR) page for details).
|
124 |
+
However 3 classes corresponding to uncertain labelisation (Mixed (16), Ligneous (17) and Other (19)) and 1 class with very poor label quantity (Clear cut (15)) were deasctivated during training.
|
125 |
+
As a result, the logits produced by the model are of size 19x1, but class 15,16,17 and 19 should appear at 0 in the logits. And labels 15,16,17 and 19 never predicted in the argmax.
|
126 |
|
127 |
|
128 |
## Bias, Risks, and Limitations
|
|
|
142 |
The user should be aware that applying the model to other type of landscapes may imply a drop in model metrics.
|
143 |
|
144 |
|
|
|
|
|
145 |
{{ bias_risks_limitations | default("[More Information Needed]", true)}}
|
146 |
|
147 |
|
|
|
156 |
|
157 |
Use the code below to get started with the model.
|
158 |
|
159 |
+
|
160 |
{{ get_started_code | default("[More Information Needed]", true)}}
|
161 |
|
162 |
## Training Details
|
|
|
196 |
|
197 |
- **Training regime:** {{ training_regime | default("[More Information Needed]", true)}} <!--fp32, fp16 mixed precision, bf16 mixed precision, bf16 non-mixed precision, fp16 non-mixed precision, fp8 mixed precision -->
|
198 |
|
199 |
+
* Model architecture: Unet (implementation from the [Segmentation Models Pytorch library](https://segmentation-modelspytorch.readthedocs.io/en/latest/docs/api.html#unet)
|
200 |
+
* Encoder : Resnet-34 pre-trained with ImageNet
|
201 |
+
* Augmentation :
|
202 |
+
* VerticalFlip(p=0.5)
|
203 |
+
* HorizontalFlip(p=0.5)
|
204 |
+
* RandomRotate90(p=0.5)
|
205 |
+
* Seed: 2022
|
206 |
+
* Batch size: 10
|
207 |
+
* Optimizer : SGD
|
208 |
+
* Learning rate : 0.02
|
209 |
+
* Class Weights :
|
210 |
+
* 1: [1, building]
|
211 |
+
* 2: [1, pervious surface]
|
212 |
+
* 3: [1, impervious surface]
|
213 |
+
* 4: [1, bare soil]
|
214 |
+
* 5: [1, water]
|
215 |
+
* 6: [1, coniferous]
|
216 |
+
* 7: [1, deciduous]
|
217 |
+
* 8: [1, brushwood]
|
218 |
+
* 9: [1, vineyard]
|
219 |
+
* 10: [1,herbaceous vegetation]
|
220 |
+
* 11: [1, agricultural land]
|
221 |
+
* 12: [1, plowed land]
|
222 |
+
* 13: [1, swimming_pool]
|
223 |
+
* 14: [1, snow]
|
224 |
+
* 15: [0, clear cut]
|
225 |
+
* 16: [0, mixed]
|
226 |
+
* 17: [0, ligneous]
|
227 |
+
* 18: [1, greenhouse]
|
228 |
+
* 19: [0, other]
|
229 |
+
|
230 |
+
|
231 |
#### Speeds, Sizes, Times [optional]
|
232 |
|
233 |
<!-- This section provides information about throughput, start/end time, checkpoint size if relevant, etc. -->
|