CharlesGaydon
commited on
Commit
•
610f7f8
1
Parent(s):
316101d
Update README.md
Browse files
README.md
CHANGED
@@ -110,34 +110,34 @@ Point clouds were preprocessed for training with point subsampling, filtering of
|
|
110 |
For inference, a preprocessing as close as possible should be used. Refer to the inference configuration file, and to the Myria3D code repository (V3.8).
|
111 |
|
112 |
#### Training Hyperparameters
|
113 |
-
|
114 |
-
|
115 |
-
|
116 |
-
|
117 |
-
|
118 |
-
|
119 |
-
|
120 |
-
|
121 |
-
|
122 |
-
|
123 |
-
|
124 |
-
|
125 |
-
|
126 |
-
|
127 |
-
|
128 |
-
|
129 |
-
|
130 |
-
|
131 |
-
|
132 |
-
|
133 |
-
|
134 |
-
|
135 |
-
|
136 |
-
|
137 |
-
|
138 |
-
|
139 |
-
|
140 |
-
|
141 |
|
142 |
#### Speeds, Sizes, Times
|
143 |
|
@@ -183,7 +183,7 @@ The following illustration gives the resulting confusion matrix :
|
|
183 |
|
184 |
### Results
|
185 |
|
186 |
-
From test patches with at least 10k points (i.e. at least 4 pts/m²), we sample without cherry-picking,
|
187 |
to match matches with the following metadata: a) URBAN, b) WATER & BRIDGE, c) OTHER_PARKING, d) BUILD_GREENHOUSE, e) HIGHSLOPE.
|
188 |
|
189 |
<div style="position: relative; text-align: center;">
|
|
|
110 |
For inference, a preprocessing as close as possible should be used. Refer to the inference configuration file, and to the Myria3D code repository (V3.8).
|
111 |
|
112 |
#### Training Hyperparameters
|
113 |
+
- Model architecture: RandLa-Net (implemented with the Pytorch-Geometric framework in [Myria3D](https://github.com/IGNF/myria3d/blob/main/myria3d/models/modules/pyg_randla_net.py))
|
114 |
+
- Augmentation :
|
115 |
+
- VerticalFlip(p=0.5)
|
116 |
+
- HorizontalFlip(p=0.5)
|
117 |
+
- Features:
|
118 |
+
- Lidar: x, y, z, echo number (1-based numbering), number of echos, reflectance (a.k.a intensity)
|
119 |
+
- Colors:
|
120 |
+
- Original: RGB + Near-Infrared (colorization from aerial images by vertical pixel-point alignement)
|
121 |
+
- Derived: average color = (R+G+B)/3 and NDVI.
|
122 |
+
- Input preprocessing:
|
123 |
+
- grid sampling: 0.25 m
|
124 |
+
- random sampling: 40,000 (if higher)
|
125 |
+
- horizontal normalization: mean xy substraction
|
126 |
+
- vertical normalization: min z substraction
|
127 |
+
- coordinates normalization: division by 25 meters
|
128 |
+
- basic occlusion model: nullify color channels if echo_number > 1
|
129 |
+
- features scaling (0-1 range):
|
130 |
+
- echo number and number of echos: division by 7
|
131 |
+
- color (r, g, b, near-infrared, average color): division by 65280 (i.e. 255*256)
|
132 |
+
- features normalization:
|
133 |
+
- reflectance: log-normalization, standardization, clipping of amplitude above 3 standard deviations.
|
134 |
+
- average color: same as reflectance.
|
135 |
+
- Batch size: 10 (x 6 GPUs)
|
136 |
+
- Number of epochs : 100 (min) - 150 (max)
|
137 |
+
- Early stopping : patience 6 and val_loss as monitor criterium
|
138 |
+
- Optimizer : Adam
|
139 |
+
- Schaeduler : mode = "min", factor = 0.5, patience = 20, cooldown = 5
|
140 |
+
- Learning rate : 0.004
|
141 |
|
142 |
#### Speeds, Sizes, Times
|
143 |
|
|
|
183 |
|
184 |
### Results
|
185 |
|
186 |
+
From test patches with at least 10k points (i.e. at least 4 pts/m²), we sample patches without cherry-picking,
|
187 |
to match matches with the following metadata: a) URBAN, b) WATER & BRIDGE, c) OTHER_PARKING, d) BUILD_GREENHOUSE, e) HIGHSLOPE.
|
188 |
|
189 |
<div style="position: relative; text-align: center;">
|