v0.8: layer separation (colour plate + ink classes) before segmentation
Browse files- .gitattributes +1 -0
- README.md +27 -9
- best_fp16.pt +2 -2
- figures/layers_brgm.png +3 -0
- training/v0.8/history.csv +9 -0
- training/v0.8/metrics.json +272 -0
- training/v0.8/separator_metrics.json +1 -0
- training/v0.8/test_metrics.csv +16 -0
.gitattributes
CHANGED
|
@@ -42,3 +42,4 @@ figures/naming_region.png filter=lfs diff=lfs merge=lfs -text
|
|
| 42 |
figures/naming_zoom.png filter=lfs diff=lfs merge=lfs -text
|
| 43 |
figures/method_geometry.png filter=lfs diff=lfs merge=lfs -text
|
| 44 |
figures/method_naming.png filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 42 |
figures/naming_zoom.png filter=lfs diff=lfs merge=lfs -text
|
| 43 |
figures/method_geometry.png filter=lfs diff=lfs merge=lfs -text
|
| 44 |
figures/method_naming.png filter=lfs diff=lfs merge=lfs -text
|
| 45 |
+
figures/layers_brgm.png filter=lfs diff=lfs merge=lfs -text
|
README.md
CHANGED
|
@@ -15,7 +15,7 @@ tags:
|
|
| 15 |
|
| 16 |
# GeoRas2Vec model – scanned geological map → GIS polygons, faults and unit names
|
| 17 |
|
| 18 |
-
**Research preview (v0.
|
| 19 |
|
| 20 |
GeoRas2Vec turns a scanned or rasterised geological map into a GeoPackage that opens directly in QGIS or ArcGIS. The GeoPackage holds three things:
|
| 21 |
|
|
@@ -32,7 +32,7 @@ This repository holds the model files. The command-line tool downloads them auto
|
|
| 32 |
## Files
|
| 33 |
| file | role |
|
| 34 |
|---|---|
|
| 35 |
-
| `best_fp16.pt` | segmentation network (U-Net++ / EfficientNet-B4, RGB + Lab): 10 pixel affinities + fault map |
|
| 36 |
| `naming_model.pkl` | scores (text read on the map, polygon) pairs |
|
| 37 |
| `legend_ranker.pkl` | ranks the candidate codes of each polygon (text evidence, leader lines, legend swatch) |
|
| 38 |
| `prop_model.pkl` | decides when a code read on the map can be propagated to polygons of the same colour |
|
|
@@ -54,14 +54,28 @@ python scripts/vectorize.py my_map.tif my_legend.csv my_map.gpkg
|
|
| 54 |
See the [GitHub README](https://github.com/PerfectWin7777/georas2vec#quick-start) for the legend CSV format.
|
| 55 |
|
| 56 |
## Results (never seen during training)
|
| 57 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 58 |
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
|
| 62 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 63 |
|
| 64 |
-
|
| 65 |
|
| 66 |

|
| 67 |
|
|
@@ -97,11 +111,15 @@ of the southern sheet, models trained on the northern sheet only.
|
|
| 97 |
on realistic renderings of official vector maps.
|
| 98 |
* On sea and ice areas the map prints many texts that are not the unit's code: there GeoRas2Vec gives the right code as first suggestion instead of guessing.
|
| 99 |
* On large maps, unnamed areas can show straight cuts at the 2048 px block edges.
|
|
|
|
|
|
|
|
|
|
| 100 |
|
| 101 |
## Versions
|
| 102 |
| version | change |
|
| 103 |
|---|---|
|
| 104 |
-
| **v0.
|
|
|
|
| 105 |
| v0.6 | English attribute table by default (`--lang fr` for French), age shown once, fragments < 200 px merged into their neighbours (3× fewer polygons), no colour propagation onto white paper / ice |
|
| 106 |
| v0.5 | two-scale mutex watershed; fault lines split at junctions; topological contours (no gap, no overlap); water detection; OCR restricted to the legend alphabet; decoding of age prefixes, Ʀ, Є and superscripts; colour propagation over the whole map; block processing of large maps |
|
| 107 |
| v0.2 | first public release (`revision="v0.2"`) |
|
|
|
|
| 15 |
|
| 16 |
# GeoRas2Vec model – scanned geological map → GIS polygons, faults and unit names
|
| 17 |
|
| 18 |
+
**Research preview (v0.8).** **Code, installation and documentation: [GeoRas2Vec on GitHub](https://github.com/PerfectWin7777/georas2vec).**
|
| 19 |
|
| 20 |
GeoRas2Vec turns a scanned or rasterised geological map into a GeoPackage that opens directly in QGIS or ArcGIS. The GeoPackage holds three things:
|
| 21 |
|
|
|
|
| 32 |
## Files
|
| 33 |
| file | role |
|
| 34 |
|---|---|
|
| 35 |
+
| `best_fp16.pt` | v0.8: layer separator (U-Net / ResNet-34: colour plate + 7 ink classes) and segmentation network (U-Net++ / EfficientNet-B4, RGB + Lab + layers): 10 pixel affinities + fault map |
|
| 36 |
| `naming_model.pkl` | scores (text read on the map, polygon) pairs |
|
| 37 |
| `legend_ranker.pkl` | ranks the candidate codes of each polygon (text evidence, leader lines, legend swatch) |
|
| 38 |
| `prop_model.pkl` | decides when a code read on the map can be propagated to polygons of the same colour |
|
|
|
|
| 54 |
See the [GitHub README](https://github.com/PerfectWin7777/georas2vec#quick-start) for the legend CSV format.
|
| 55 |
|
| 56 |
## Results (never seen during training)
|
| 57 |
+
| test (large units / small units / area in the correct polygon / fault F1) | v0.7 | **v0.8** |
|
| 58 |
+
|---|---|---|
|
| 59 |
+
| real scans, USGS GMNA 300 dpi (78 tiles, official test split) | **94.5** / 77.0 / **98.5** / 0.78 | 93.0 / **77.4** / 98.0 / 0.78 |
|
| 60 |
+
| same scans with a topographic background printed over them | 85.5 / 73.0 / 73.7 / 0.44 | **89.1** / **75.2** / **96.9** / **0.78** |
|
| 61 |
+
| BRGM 1:50 000 sheet, Calvados (map series never seen in training) | **83.1** / **63.4** / 70.4 / 0.07 | 82.9 / 42.7 / **90.0** / **0.45** |
|
| 62 |
+
| synthetic maps, 4 continents (530 tiles) | 87.5 / 76.2 / 91.2 / 0.98 | **88.0** / **76.6** / **91.8** / 0.97 |
|
| 63 |
|
| 64 |
+
<sub>IoU ≥ 0.5; small = < 1000 px; fault F1 at 2 px. Protocols and every training run, including the failed
|
| 65 |
+
ones: [docs/EXPERIMENTS.md on GitHub](https://github.com/PerfectWin7777/georas2vec/blob/main/docs/EXPERIMENTS.md).</sub>
|
| 66 |
+
|
| 67 |
+
### New in v0.8: maps printed over a topographic base
|
| 68 |
+
|
| 69 |
+
Detailed maps (e.g. French 1:50 000 sheets) print the geology over contour lines, roads and field boundaries. Up to
|
| 70 |
+
v0.7 the network relied on the shortcut *thin black line = unit boundary*, true on small-scale maps but false on
|
| 71 |
+
detailed ones. Training one network on both kinds of maps cost 2 to 4 points on small-scale scans, even with
|
| 72 |
+
rebalanced data or a network conditioned on the map style. v0.8 instead **separates the printed layers first**:
|
| 73 |
+
a separator predicts the colour plate (map without ink) and the kind of ink on every pixel, and the segmentation
|
| 74 |
+
network reads the image together with these layers. It starts from v0.7 weights. On maps with a topographic
|
| 75 |
+
background the area in the correct polygon rises from 73.7 % to 96.9 %; the price is 1.5 points on large units of
|
| 76 |
+
clean scans. Use `revision="v0.7"` for the previous model.
|
| 77 |
|
| 78 |
+

|
| 79 |
|
| 80 |

|
| 81 |
|
|
|
|
| 111 |
on realistic renderings of official vector maps.
|
| 112 |
* On sea and ice areas the map prints many texts that are not the unit's code: there GeoRas2Vec gives the right code as first suggestion instead of guessing.
|
| 113 |
* On large maps, unnamed areas can show straight cuts at the 2048 px block edges.
|
| 114 |
+
* Detailed maps with a topographic background: small units are often merged (42.7 % recovered on the French test
|
| 115 |
+
sheet vs 63.4 % with v0.7), and the separator still labels many roads as faults on real sheets.
|
| 116 |
+
* The naming results above were measured with the v0.7 segmentation network.
|
| 117 |
|
| 118 |
## Versions
|
| 119 |
| version | change |
|
| 120 |
|---|---|
|
| 121 |
+
| **v0.8** | layer separation before segmentation (colour plate + 7 ink classes): robust to topographic base maps, faults detected on French 1:50 000 sheets |
|
| 122 |
+
| v0.7 | automatic legend reading (swatch boxes + codes), geological code reader (56/60 real legend boxes vs 26/60 with generic OCR), learned swatch comparator, ranker trained with real printed legends |
|
| 123 |
| v0.6 | English attribute table by default (`--lang fr` for French), age shown once, fragments < 200 px merged into their neighbours (3× fewer polygons), no colour propagation onto white paper / ice |
|
| 124 |
| v0.5 | two-scale mutex watershed; fault lines split at junctions; topological contours (no gap, no overlap); water detection; OCR restricted to the legend alphabet; decoding of age prefixes, Ʀ, Є and superscripts; colour propagation over the whole map; block processing of large maps |
|
| 125 |
| v0.2 | first public release (`revision="v0.2"`) |
|
best_fp16.pt
CHANGED
|
@@ -1,3 +1,3 @@
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
-
oid sha256:
|
| 3 |
-
size
|
|
|
|
| 1 |
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:2372a44946bbff7a50035472063d2511fd4a5baa9ec68fe44305c9bdf170d1c5
|
| 3 |
+
size 91172097
|
figures/layers_brgm.png
ADDED
|
Git LFS Details
|
training/v0.8/history.csv
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
step,loss,heures,modele_score,modele_reel,modele_synth,meilleur,ema_score,ema_reel,ema_synth
|
| 2 |
+
2500,0.7154660440564156,1.24,0.08204887501762081,0.05205817378573897,0.11203957624950267,ema,0.07850958571485984,0.04677064077515888,0.1102485306545608
|
| 3 |
+
5000,0.691128010737896,1.79,0.07865927065232332,0.04792626898096005,0.10939227232368659,,0.07871562792559383,0.0468937479860099,0.11053750786517776
|
| 4 |
+
7500,0.6835786717057228,2.34,0.07597521601377022,0.05008050776863225,0.10186992425890817,modele,0.07959509787023923,0.04946425209015247,0.109725943650326
|
| 5 |
+
10000,0.6747275333285332,2.88,0.07450312302617934,0.04659467867239357,0.10241156737996511,modele,0.07743297435999716,0.04743463793002789,0.10743131078996644
|
| 6 |
+
12500,0.6710859683036804,3.43,0.07746846829117007,0.049868209859521194,0.10506872672281896,,0.07550114371605977,0.04828007367063503,0.10272221376148451
|
| 7 |
+
15000,0.6661567495465278,3.98,0.07771220143283196,0.04933374534695209,0.10609065751871183,,0.07837112128068843,0.05050647629489115,0.10623576626648572
|
| 8 |
+
17500,0.6586158741235733,4.52,0.07595049465585083,0.04855999935642912,0.10334098995527254,,0.07640218282653927,0.048853154051914226,0.10395121160116431
|
| 9 |
+
20000,0.6567949496746063,5.07,0.07547126947633316,0.04850230238691251,0.10244023656575381,,0.07539654798573334,0.04846927762533615,0.10232381834613054
|
training/v0.8/metrics.json
ADDED
|
@@ -0,0 +1,272 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cfg": {
|
| 3 |
+
"arch": "unetpp",
|
| 4 |
+
"encoder": "efficientnet-b4",
|
| 5 |
+
"input": "rgb+lab",
|
| 6 |
+
"offsets": [
|
| 7 |
+
[
|
| 8 |
+
1,
|
| 9 |
+
0
|
| 10 |
+
],
|
| 11 |
+
[
|
| 12 |
+
0,
|
| 13 |
+
1
|
| 14 |
+
],
|
| 15 |
+
[
|
| 16 |
+
3,
|
| 17 |
+
0
|
| 18 |
+
],
|
| 19 |
+
[
|
| 20 |
+
0,
|
| 21 |
+
3
|
| 22 |
+
],
|
| 23 |
+
[
|
| 24 |
+
9,
|
| 25 |
+
0
|
| 26 |
+
],
|
| 27 |
+
[
|
| 28 |
+
0,
|
| 29 |
+
9
|
| 30 |
+
],
|
| 31 |
+
[
|
| 32 |
+
9,
|
| 33 |
+
9
|
| 34 |
+
],
|
| 35 |
+
[
|
| 36 |
+
9,
|
| 37 |
+
-9
|
| 38 |
+
],
|
| 39 |
+
[
|
| 40 |
+
27,
|
| 41 |
+
0
|
| 42 |
+
],
|
| 43 |
+
[
|
| 44 |
+
0,
|
| 45 |
+
27
|
| 46 |
+
]
|
| 47 |
+
],
|
| 48 |
+
"n_short": 2,
|
| 49 |
+
"mws_stride": 1,
|
| 50 |
+
"mws_min_size": 40,
|
| 51 |
+
"tile": 512,
|
| 52 |
+
"sep": true,
|
| 53 |
+
"small_w": false,
|
| 54 |
+
"ohem": false,
|
| 55 |
+
"tta": false
|
| 56 |
+
},
|
| 57 |
+
"steps": 10000,
|
| 58 |
+
"val_score": 0.07450312302617934,
|
| 59 |
+
"test": [
|
| 60 |
+
{
|
| 61 |
+
"groupe": "Africa",
|
| 62 |
+
"tuiles": 70,
|
| 63 |
+
"ARAND": 0.0277,
|
| 64 |
+
"sur_decoupage": 0.1428,
|
| 65 |
+
"fusion": 0.1006,
|
| 66 |
+
"rappel_petites": 0.8858,
|
| 67 |
+
"rappel_grandes": 0.9823,
|
| 68 |
+
"rappel_40_200": 0.8,
|
| 69 |
+
"rappel_200_1000": 0.9321,
|
| 70 |
+
"rappel_tous": 0.945,
|
| 71 |
+
"surface_correcte": 0.9933,
|
| 72 |
+
"F1_failles": NaN
|
| 73 |
+
},
|
| 74 |
+
{
|
| 75 |
+
"groupe": "Europe",
|
| 76 |
+
"tuiles": 184,
|
| 77 |
+
"ARAND": 0.1035,
|
| 78 |
+
"sur_decoupage": 0.4981,
|
| 79 |
+
"fusion": 0.3232,
|
| 80 |
+
"rappel_petites": 0.8187,
|
| 81 |
+
"rappel_grandes": 0.8905,
|
| 82 |
+
"rappel_40_200": 0.7788,
|
| 83 |
+
"rappel_200_1000": 0.8851,
|
| 84 |
+
"rappel_tous": 0.8363,
|
| 85 |
+
"surface_correcte": 0.9233,
|
| 86 |
+
"F1_failles": 0.0243
|
| 87 |
+
},
|
| 88 |
+
{
|
| 89 |
+
"groupe": "North America",
|
| 90 |
+
"tuiles": 268,
|
| 91 |
+
"ARAND": 0.09,
|
| 92 |
+
"sur_decoupage": 0.4352,
|
| 93 |
+
"fusion": 0.2824,
|
| 94 |
+
"rappel_petites": 0.7289,
|
| 95 |
+
"rappel_grandes": 0.9029,
|
| 96 |
+
"rappel_40_200": 0.6656,
|
| 97 |
+
"rappel_200_1000": 0.797,
|
| 98 |
+
"rappel_tous": 0.7702,
|
| 99 |
+
"surface_correcte": 0.95,
|
| 100 |
+
"F1_failles": 0.9473
|
| 101 |
+
},
|
| 102 |
+
{
|
| 103 |
+
"groupe": "Oceania",
|
| 104 |
+
"tuiles": 150,
|
| 105 |
+
"ARAND": 0.171,
|
| 106 |
+
"sur_decoupage": 0.9302,
|
| 107 |
+
"fusion": 0.4689,
|
| 108 |
+
"rappel_petites": 0.6476,
|
| 109 |
+
"rappel_grandes": 0.765,
|
| 110 |
+
"rappel_40_200": 0.6163,
|
| 111 |
+
"rappel_200_1000": 0.6949,
|
| 112 |
+
"rappel_tous": 0.6657,
|
| 113 |
+
"surface_correcte": 0.8421,
|
| 114 |
+
"F1_failles": 0.9871
|
| 115 |
+
},
|
| 116 |
+
{
|
| 117 |
+
"groupe": "BRGM-50k / real",
|
| 118 |
+
"tuiles": 64,
|
| 119 |
+
"ARAND": 0.1231,
|
| 120 |
+
"sur_decoupage": 0.3039,
|
| 121 |
+
"fusion": 0.422,
|
| 122 |
+
"rappel_petites": 0.4577,
|
| 123 |
+
"rappel_grandes": 0.8234,
|
| 124 |
+
"rappel_40_200": 0.2041,
|
| 125 |
+
"rappel_200_1000": 0.5914,
|
| 126 |
+
"rappel_tous": 0.7624,
|
| 127 |
+
"surface_correcte": 0.9059,
|
| 128 |
+
"F1_failles": 0.0243
|
| 129 |
+
},
|
| 130 |
+
{
|
| 131 |
+
"groupe": "GA-Australia / synthetic",
|
| 132 |
+
"tuiles": 150,
|
| 133 |
+
"ARAND": 0.171,
|
| 134 |
+
"sur_decoupage": 0.9302,
|
| 135 |
+
"fusion": 0.4689,
|
| 136 |
+
"rappel_petites": 0.6476,
|
| 137 |
+
"rappel_grandes": 0.765,
|
| 138 |
+
"rappel_40_200": 0.6163,
|
| 139 |
+
"rappel_200_1000": 0.6949,
|
| 140 |
+
"rappel_tous": 0.6657,
|
| 141 |
+
"surface_correcte": 0.8421,
|
| 142 |
+
"F1_failles": 0.9871
|
| 143 |
+
},
|
| 144 |
+
{
|
| 145 |
+
"groupe": "GEO-LiM / synthetic",
|
| 146 |
+
"tuiles": 120,
|
| 147 |
+
"ARAND": 0.0931,
|
| 148 |
+
"sur_decoupage": 0.6017,
|
| 149 |
+
"fusion": 0.2705,
|
| 150 |
+
"rappel_petites": 0.8245,
|
| 151 |
+
"rappel_grandes": 0.9123,
|
| 152 |
+
"rappel_40_200": 0.7839,
|
| 153 |
+
"rappel_200_1000": 0.8935,
|
| 154 |
+
"rappel_tous": 0.842,
|
| 155 |
+
"surface_correcte": 0.9332,
|
| 156 |
+
"F1_failles": NaN
|
| 157 |
+
},
|
| 158 |
+
{
|
| 159 |
+
"groupe": "GMNA / real",
|
| 160 |
+
"tuiles": 78,
|
| 161 |
+
"ARAND": 0.0535,
|
| 162 |
+
"sur_decoupage": 0.225,
|
| 163 |
+
"fusion": 0.2433,
|
| 164 |
+
"rappel_petites": 0.7484,
|
| 165 |
+
"rappel_grandes": 0.9124,
|
| 166 |
+
"rappel_40_200": 0.5242,
|
| 167 |
+
"rappel_200_1000": 0.8539,
|
| 168 |
+
"rappel_tous": 0.8245,
|
| 169 |
+
"surface_correcte": 0.9758,
|
| 170 |
+
"F1_failles": 0.6355
|
| 171 |
+
},
|
| 172 |
+
{
|
| 173 |
+
"groupe": "GMNA / synthetic",
|
| 174 |
+
"tuiles": 60,
|
| 175 |
+
"ARAND": 0.0223,
|
| 176 |
+
"sur_decoupage": 0.1294,
|
| 177 |
+
"fusion": 0.1246,
|
| 178 |
+
"rappel_petites": 0.8943,
|
| 179 |
+
"rappel_grandes": 0.9739,
|
| 180 |
+
"rappel_40_200": 0.8065,
|
| 181 |
+
"rappel_200_1000": 0.9366,
|
| 182 |
+
"rappel_tous": 0.929,
|
| 183 |
+
"surface_correcte": 0.9909,
|
| 184 |
+
"F1_failles": 0.9708
|
| 185 |
+
},
|
| 186 |
+
{
|
| 187 |
+
"groupe": "UNESCO-Ethiopia / synthetic",
|
| 188 |
+
"tuiles": 70,
|
| 189 |
+
"ARAND": 0.0277,
|
| 190 |
+
"sur_decoupage": 0.1428,
|
| 191 |
+
"fusion": 0.1006,
|
| 192 |
+
"rappel_petites": 0.8858,
|
| 193 |
+
"rappel_grandes": 0.9823,
|
| 194 |
+
"rappel_40_200": 0.8,
|
| 195 |
+
"rappel_200_1000": 0.9321,
|
| 196 |
+
"rappel_tous": 0.945,
|
| 197 |
+
"surface_correcte": 0.9933,
|
| 198 |
+
"F1_failles": NaN
|
| 199 |
+
},
|
| 200 |
+
{
|
| 201 |
+
"groupe": "USGS-California / synthetic",
|
| 202 |
+
"tuiles": 70,
|
| 203 |
+
"ARAND": 0.0761,
|
| 204 |
+
"sur_decoupage": 0.5069,
|
| 205 |
+
"fusion": 0.3483,
|
| 206 |
+
"rappel_petites": 0.7503,
|
| 207 |
+
"rappel_grandes": 0.8986,
|
| 208 |
+
"rappel_40_200": 0.6966,
|
| 209 |
+
"rappel_200_1000": 0.7996,
|
| 210 |
+
"rappel_tous": 0.7841,
|
| 211 |
+
"surface_correcte": 0.9489,
|
| 212 |
+
"F1_failles": 0.9689
|
| 213 |
+
},
|
| 214 |
+
{
|
| 215 |
+
"groupe": "USGS-Nevada / synthetic",
|
| 216 |
+
"tuiles": 60,
|
| 217 |
+
"ARAND": 0.2212,
|
| 218 |
+
"sur_decoupage": 0.9305,
|
| 219 |
+
"fusion": 0.4142,
|
| 220 |
+
"rappel_petites": 0.6914,
|
| 221 |
+
"rappel_grandes": 0.8679,
|
| 222 |
+
"rappel_40_200": 0.6483,
|
| 223 |
+
"rappel_200_1000": 0.7544,
|
| 224 |
+
"rappel_tous": 0.7204,
|
| 225 |
+
"surface_correcte": 0.8795,
|
| 226 |
+
"F1_failles": 0.9468
|
| 227 |
+
},
|
| 228 |
+
{
|
| 229 |
+
"groupe": "TEST COMPLET",
|
| 230 |
+
"tuiles": 672,
|
| 231 |
+
"ARAND": 0.1053,
|
| 232 |
+
"sur_decoupage": 0.5324,
|
| 233 |
+
"fusion": 0.3163,
|
| 234 |
+
"rappel_petites": 0.7096,
|
| 235 |
+
"rappel_grandes": 0.8632,
|
| 236 |
+
"rappel_40_200": 0.6645,
|
| 237 |
+
"rappel_200_1000": 0.7699,
|
| 238 |
+
"rappel_tous": 0.7421,
|
| 239 |
+
"surface_correcte": 0.9229,
|
| 240 |
+
"F1_failles": 0.9624
|
| 241 |
+
},
|
| 242 |
+
{
|
| 243 |
+
"groupe": "PIÈGE unités de même couleur",
|
| 244 |
+
"tuiles": 48,
|
| 245 |
+
"ARAND": 0.164,
|
| 246 |
+
"sur_decoupage": 0.5749,
|
| 247 |
+
"fusion": 0.5494,
|
| 248 |
+
"rappel_petites": 0.6101,
|
| 249 |
+
"rappel_grandes": 0.741,
|
| 250 |
+
"rappel_40_200": 0.5677,
|
| 251 |
+
"rappel_200_1000": 0.6596,
|
| 252 |
+
"rappel_tous": 0.6404,
|
| 253 |
+
"surface_correcte": 0.8339,
|
| 254 |
+
"F1_failles": 0.9711
|
| 255 |
+
},
|
| 256 |
+
{
|
| 257 |
+
"groupe": "PIÈGE unités à motifs",
|
| 258 |
+
"tuiles": 48,
|
| 259 |
+
"ARAND": 0.0929,
|
| 260 |
+
"sur_decoupage": 0.48,
|
| 261 |
+
"fusion": 0.2884,
|
| 262 |
+
"rappel_petites": 0.7466,
|
| 263 |
+
"rappel_grandes": 0.901,
|
| 264 |
+
"rappel_40_200": 0.6976,
|
| 265 |
+
"rappel_200_1000": 0.8011,
|
| 266 |
+
"rappel_tous": 0.7834,
|
| 267 |
+
"surface_correcte": 0.9435,
|
| 268 |
+
"F1_failles": 0.9771
|
| 269 |
+
}
|
| 270 |
+
],
|
| 271 |
+
"train_hours": 5.23
|
| 272 |
+
}
|
training/v0.8/separator_metrics.json
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"plate_err_raw": 81.95002736141336, "plate_err": 13.90600299835205, "iou": [0.8818230283938705, 0.5564695664286939, 0.318674142871791, 0.6345497138853331, 0.4833100426534778, 0.21626427395051015, 0.6380145011186716]}
|
training/v0.8/test_metrics.csv
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
groupe,tuiles,ARAND,sur_decoupage,fusion,rappel_petites,rappel_grandes,rappel_40_200,rappel_200_1000,rappel_tous,surface_correcte,F1_failles
|
| 2 |
+
Africa,70,0.0277,0.1428,0.1006,0.8858,0.9823,0.8,0.9321,0.945,0.9933,
|
| 3 |
+
Europe,184,0.1035,0.4981,0.3232,0.8187,0.8905,0.7788,0.8851,0.8363,0.9233,0.0243
|
| 4 |
+
North America,268,0.09,0.4352,0.2824,0.7289,0.9029,0.6656,0.797,0.7702,0.95,0.9473
|
| 5 |
+
Oceania,150,0.171,0.9302,0.4689,0.6476,0.765,0.6163,0.6949,0.6657,0.8421,0.9871
|
| 6 |
+
BRGM-50k / real,64,0.1231,0.3039,0.422,0.4577,0.8234,0.2041,0.5914,0.7624,0.9059,0.0243
|
| 7 |
+
GA-Australia / synthetic,150,0.171,0.9302,0.4689,0.6476,0.765,0.6163,0.6949,0.6657,0.8421,0.9871
|
| 8 |
+
GEO-LiM / synthetic,120,0.0931,0.6017,0.2705,0.8245,0.9123,0.7839,0.8935,0.842,0.9332,
|
| 9 |
+
GMNA / real,78,0.0535,0.225,0.2433,0.7484,0.9124,0.5242,0.8539,0.8245,0.9758,0.6355
|
| 10 |
+
GMNA / synthetic,60,0.0223,0.1294,0.1246,0.8943,0.9739,0.8065,0.9366,0.929,0.9909,0.9708
|
| 11 |
+
UNESCO-Ethiopia / synthetic,70,0.0277,0.1428,0.1006,0.8858,0.9823,0.8,0.9321,0.945,0.9933,
|
| 12 |
+
USGS-California / synthetic,70,0.0761,0.5069,0.3483,0.7503,0.8986,0.6966,0.7996,0.7841,0.9489,0.9689
|
| 13 |
+
USGS-Nevada / synthetic,60,0.2212,0.9305,0.4142,0.6914,0.8679,0.6483,0.7544,0.7204,0.8795,0.9468
|
| 14 |
+
TEST COMPLET,672,0.1053,0.5324,0.3163,0.7096,0.8632,0.6645,0.7699,0.7421,0.9229,0.9624
|
| 15 |
+
PIÈGE unités de même couleur,48,0.164,0.5749,0.5494,0.6101,0.741,0.5677,0.6596,0.6404,0.8339,0.9711
|
| 16 |
+
PIÈGE unités à motifs,48,0.0929,0.48,0.2884,0.7466,0.901,0.6976,0.8011,0.7834,0.9435,0.9771
|