Skip to content

Commit

Permalink
Merge pull request #7 from swiss-territorial-data-lab/ac/rdp
Browse files Browse the repository at this point in the history 
Simplify polygons associated with predictions with the Ramer-Douglas-Peucker algorithm
  • Loading branch information
@cleherny
cleherny authored Jan 27, 2023
2 parents af81652 + 90854d6 commit 91f00d0
Show file tree
Hide file tree
Showing 9 changed files with 110 additions and 157 deletions.
12 changes: 8 additions & 4 deletions examples/swimming-pool-detection/GE/config_GE.yaml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -57,6 +57,10 @@ make_predictions.py:
detectron2_config_file: '../detectron2_config_GE.yaml' # path relative to the working_folder
model_weights:
pth_file: 'logs/model_final.pth'
image_metadata_json: output_GE/img_metadata.json
rdp_simplification: # rdp = Ramer-Douglas-Peucker
enabled: true
epsilon: 0.5 # cf. https://rdp.readthedocs.io/en/latest/
score_lower_threshold: 0.05

assess_predictions.py:
Expand All @@ -66,8 +70,8 @@ assess_predictions.py:
image_metadata_json: output_GE/img_metadata.json
split_aoi_tiles_geojson: output_GE/split_aoi_tiles.geojson # aoi = Area of Interest
predictions:
trn: output_GE/trn_predictions_at_0dot05_threshold.pkl
val: output_GE/val_predictions_at_0dot05_threshold.pkl
tst: output_GE/tst_predictions_at_0dot05_threshold.pkl
oth: output_GE/oth_predictions_at_0dot05_threshold.pkl
trn: output_GE/trn_predictions_at_0dot05_threshold.gpkg
val: output_GE/val_predictions_at_0dot05_threshold.gpkg
tst: output_GE/tst_predictions_at_0dot05_threshold.gpkg
oth: output_GE/oth_predictions_at_0dot05_threshold.gpkg
output_folder: output_GE
14 changes: 9 additions & 5 deletions examples/swimming-pool-detection/NE/config_NE.yaml
View file
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,7 @@ train_model.py:
detectron2_config_file: '../detectron2_config_NE.yaml' # path relative to the working_folder
model_weights:
model_zoo_checkpoint_url: "COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml"

make_predictions.py:
working_folder: output_NE
log_subfolder: logs
Expand All @@ -59,6 +59,10 @@ make_predictions.py:
detectron2_config_file: '../detectron2_config_NE.yaml' # path relative to the working_folder
model_weights:
pth_file: './logs/model_final.pth'
image_metadata_json: output_NE/img_metadata.json
rdp_simplification: # rdp = Ramer-Douglas-Peucker
enabled: true
epsilon: 0.5 # cf. https://rdp.readthedocs.io/en/latest/
score_lower_threshold: 0.05

assess_predictions.py:
Expand All @@ -68,8 +72,8 @@ assess_predictions.py:
image_metadata_json: output_NE/img_metadata.json
split_aoi_tiles_geojson: output_NE/split_aoi_tiles.geojson # aoi = Area of Interest
predictions:
trn: output_NE/trn_predictions_at_0dot05_threshold.pkl
val: output_NE/val_predictions_at_0dot05_threshold.pkl
tst: output_NE/tst_predictions_at_0dot05_threshold.pkl
oth: output_NE/oth_predictions_at_0dot05_threshold.pkl
trn: output_NE/trn_predictions_at_0dot05_threshold.gpkg
val: output_NE/val_predictions_at_0dot05_threshold.gpkg
tst: output_NE/tst_predictions_at_0dot05_threshold.gpkg
oth: output_NE/oth_predictions_at_0dot05_threshold.gpkg
output_folder: output_NE
50 changes: 41 additions & 9 deletions helpers/detectron2.py
View file
Original file line number Diff line number Diff line change
@@ -1,7 +1,7 @@
#!/usr/bin/env python
# coding: utf-8

import os
import os, sys
import time
import torch
import numpy as np
Expand All @@ -16,6 +16,11 @@
from detectron2.utils import comm
from detectron2.utils.logger import log_every_n_seconds

from rasterio import features
from shapely.affinity import affine_transform
from shapely.geometry import box
from rdp import rdp

# cf. https://medium.com/@apofeniaco/training-on-detectron2-with-a-validation-set-and-plot-loss-on-it-to-avoid-overfitting-6449418fbf4e
# cf. https://towardsdatascience.com/face-detection-on-custom-dataset-with-detectron2-and-pytorch-using-python-23c17e99e162
# cf. http://cocodataset.org/#detection-eval
Expand Down Expand Up @@ -143,23 +148,50 @@ def _preprocess(preds):
return out


def dt2predictions_to_list(preds):
def detectron2preds_to_features(preds, crs, transform, rdp_enabled, rdp_eps):

instances = []
feats = []

tmp = _preprocess(preds)

for idx in range(len(tmp['scores'])):

instance = {}
instance['score'] = tmp['scores'][idx]
instance['pred_class'] = tmp['pred_classes'][idx]

if 'pred_masks' in tmp.keys():
instance['pred_mask'] = tmp['pred_masks'][idx]

instance['pred_box'] = tmp['pred_boxes'][idx]

instances.append(instance)

return instances
pred_mask_int = tmp['pred_masks'][idx].astype(np.uint8)
_feats = [
{
'type': 'Feature',
'properties': {'score': instance['score'], 'crs': crs},
'geometry': geom
} for (geom, v) in features.shapes(pred_mask_int, mask=None, transform=transform) if v == 1.0
]

for f in _feats:
if rdp_enabled:
coords = f['geometry']['coordinates']
coords_after_rdp = [rdp(x, epsilon=rdp_eps) for x in coords]
f['geometry']['coordinates'] = coords_after_rdp

feats.append(f)

else: # if pred_masks does not exist, pred_boxes should (it depends on Detectron2's MASK_ON config param)
instance['pred_box'] = tmp['pred_boxes'][idx]

geom = affine_transform(box(*instance['pred_box']), [transform.a, transform.b, transform.d, transform.e, transform.xoff, transform.yoff])
_feats = [
{
'type': 'Feature',
'properties': {'score': instance['score'], 'crs': crs},
'geometry': geom
}
]

feats += _feats

return feats

88 changes: 1 addition & 87 deletions helpers/misc.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -36,75 +36,6 @@ def my_unpack(list_of_tuples):
return [item for t in list_of_tuples for item in t]


# cf. https://gis.stackexchange.com/questions/187877/how-to-polygonize-raster-to-shapely-polygons
def predictions_to_features(predictions_dict, img_path):
"""
predictions_dict = {"<image_filename>': [<prediction>]
<prediction> = {'score': ..., 'pred_class': ..., 'pred_mask': ..., 'pred_box': ...}
"""

feats = []

for k, v in predictions_dict.items():
# N.B.: src images are only used for georeferencing (src.crs, src.transform)
with rasterio.open(os.path.join(img_path, k)) as src:

for pred in v:

pred_mask_int = pred['pred_mask'].astype(int)

feats += [{'type': 'Feature',
'properties': {'raster_val': v, 'score': pred['score'], 'crs': src.crs},
'geometry': s
} for (s, v) in features.shapes(pred_mask_int, mask=None, transform=src.transform)
]

return feats


def fast_predictions_to_features(predictions_dict, img_metadata_dict):
"""
predictions_dict = {"<image_filename>': [<prediction>]
<prediction> = {'score': ..., 'pred_class': ..., 'pred_mask': ..., 'pred_box': ...}
img_metadata_dict's values includes the metadata issued by ArcGIS Server; keys are equal to filenames
"""

feats = []

for k, v in predictions_dict.items():

# k is like "images/val-images-256/18_135617_92947.tif"
# img_metadata_dict keys are like "18_135617_92947.tif"

kk = k.split('/')[-1]
this_img_metadata = img_metadata_dict[kk]
#print(this_img_metadata)

crs = f"EPSG:{this_img_metadata['extent']['spatialReference']['latestWkid']}"
transform = image_metadata_to_affine_transform(this_img_metadata)
#print(transform)
for pred in v:
#print(pred)
if 'pred_mask' in pred.keys():

pred_mask_int = pred['pred_mask'].astype(np.uint8)
feats += [{'type': 'Feature',
'properties': {'raster_val': v, 'score': pred['score'], 'crs': crs},
'geometry': s
} for (s, v) in features.shapes(pred_mask_int, mask=None, transform=transform)
]

else:

geom = affine_transform(box(*pred['pred_box']), [transform.a, transform.b, transform.d, transform.e, transform.xoff, transform.yoff])
feats += [{'type': 'Feature',
'properties': {'raster_val': 1.0, 'score': pred['score'], 'crs': crs},
'geometry': geom}]

return feats


def img_md_record_to_tile_id(img_md_record):

filename = os.path.split(img_md_record.img_file)[-1]
Expand Down Expand Up @@ -291,21 +222,4 @@ def image_metadata_to_world_file(image_metadata):
c += a/2.0 # <- IMPORTANT
f += e/2.0 # <- IMPORTANT

return "\n".join([str(a), str(d), str(b), str(e), str(c), str(f)+"\n"])


def image_metadata_to_affine_transform(image_metadata):
"""
This uses rasterio.
"""

xmin = image_metadata['extent']['xmin']
xmax = image_metadata['extent']['xmax']
ymin = image_metadata['extent']['ymin']
ymax = image_metadata['extent']['ymax']
width = image_metadata['width']
height = image_metadata['height']

affine = from_bounds(xmin, ymin, xmax, ymax, width, height)

return affine
return "\n".join([str(a), str(d), str(b), str(e), str(c), str(f)+"\n"])
1 change: 1 addition & 0 deletions requirements.in
View file
Original file line number Diff line number Diff line change
Expand Up @@ -18,3 +18,4 @@ torch @ https://download.pytorch.org/whl/cu113/torch-1.10.2%2Bcu113-cp38-cp38-li
torchvision @ https://download.pytorch.org/whl/cu113/torchvision-0.11.3%2Bcu113-cp38-cp38-linux_x86_64.whl
detectron2 @ https://dl.fbaipublicfiles.com/detectron2/wheels/cu113/torch1.10/detectron2-0.6%2Bcu113-cp38-cp38-linux_x86_64.whl
plotly
rdp
3 changes: 3 additions & 0 deletions requirements.txt
View file
Original file line number Diff line number Diff line change
Expand Up @@ -118,6 +118,7 @@ numpy==1.23.3
# pandas
# pycocotools
# rasterio
# rdp
# snuggs
# supermercado
# tensorboard
Expand Down Expand Up @@ -186,6 +187,8 @@ rasterio==1.3.2
# via
# -r requirements.in
# supermercado
rdp==0.8
# via -r requirements.in
regex==2022.9.13
# via black
requests==2.28.1
Expand Down
40 changes: 5 additions & 35 deletions scripts/assess_predictions.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -122,47 +122,19 @@

# ------ Loading predictions

preds_dict = {}

for dataset, preds_file in PREDICTION_FILES.items():
with open(preds_file, 'rb') as fp:
preds_dict[dataset] = pickle.load(fp)

# ------ Extracting vector features out of predictions

preds_gdf_dict = {}

logger.info(f'Extracting vector features...')
tic = time.time()
tqdm_log = tqdm(total=len(preds_dict.keys()), position=0)

for dataset, preds in preds_dict.items():

tqdm_log.set_description_str(f'Current dataset: {dataset}')

features = misc.fast_predictions_to_features(preds, img_metadata_dict=img_metadata_dict)
gdf = gpd.GeoDataFrame.from_features(features)
gdf['dataset'] = dataset
gdf.crs = features[0]['properties']['crs']

preds_gdf_dict[dataset] = gdf[gdf.raster_val == 1.0][['geometry', 'score', 'dataset']]

file_to_write = os.path.join(OUTPUT_DIR, f"{dataset}_predictions.geojson")
preds_gdf_dict[dataset].to_crs(epsg=4326).to_file(file_to_write, driver='GeoJSON', index=False)
written_files.append(file_to_write)

tqdm_log.update(1)
for dataset, preds_file in PREDICTION_FILES.items():
preds_gdf_dict[dataset] = gpd.read_file(preds_file)

tqdm_log.close()
logger.info(f'...done. Elapsed time = {(time.time()-tic):.2f} seconds.')

if len(labels_gdf)>0:

# ------ Comparing predictions with ground-truth data and computing metrics

# init
metrics = {}
for dataset in preds_dict.keys():
for dataset in preds_gdf_dict.keys():
metrics[dataset] = []

metrics_df_dict = {}
Expand Down Expand Up @@ -280,8 +252,6 @@
written_files.append(file_to_write)




# ------ tagging predictions

# we select the threshold which maximizes the f1-score on the val dataset
Expand Down Expand Up @@ -315,8 +285,8 @@
tagged_preds_gdf_dict[x] for x in metrics.keys()
])

file_to_write = os.path.join(OUTPUT_DIR, f'tagged_predictions.geojson')
tagged_preds_gdf[['geometry', 'score', 'tag', 'dataset']].to_crs(epsg=4326).to_file(file_to_write, driver='GeoJSON', index=False)
file_to_write = os.path.join(OUTPUT_DIR, f'tagged_predictions.gpkg')
tagged_preds_gdf[['geometry', 'score', 'tag', 'dataset']].to_file(file_to_write, driver='GPKG', index=False)
written_files.append(file_to_write)

# ------ wrap-up
Expand Down
Loading

0 comments on commit 91f00d0

Please sign in to comment.