def _get_detections(generator,
model,
score_threshold=0.05,
max_detections=100,
save_path=None,
experiment=None):
""" Get the detections from the model using the generator.
The result is a list of lists such that the size is:
all_detections[num_images][num_classes] = detections[num_detections, 4 + num_classes]
# Arguments
generator : The generator used to run images through the model.
model : The model to run on the images.
score_threshold : The score confidence threshold to use.
max_detections : The maximum number of detections to use per image.
save_path : The path to save the images with visualized detections to.
experiment : Comet ML experiment
# Returns
A list of lists containing the detections for each image in the generator.
"""
all_detections = [[None for i in range(generator.num_classes())]
for j in range(generator.size())]
for i in range(generator.size()):
raw_image = generator.load_image(i)
#need to make contigious see https://stackoverflow.com/questions/23830618/python-opencv-typeerror-layout-of-the-output-array-incompatible-with-cvmat
raw_image = raw_image.copy()
#Skip if missing a component data source
if raw_image is None:
print("Empty image, skipping")
continue
image = generator.preprocess_image(raw_image)
image, scale = generator.resize_image(image)
if keras.backend.image_data_format() == 'channels_first':
image = image.transpose((2, 0, 1))
# run network
boxes, scores, labels = model.predict_on_batch(
np.expand_dims(image, axis=0))[:3]
# correct boxes for image scale
boxes /= scale
# select indices which have a score above the threshold
indices = np.where(scores[0, :] > score_threshold)[0]
# select those scores
scores = scores[0][indices]
# find the order with which to sort the scores
scores_sort = np.argsort(-scores)[:max_detections]
# select detections
image_boxes = boxes[0, indices[scores_sort], :]
image_scores = scores[scores_sort]
image_labels = labels[0, indices[scores_sort]]
image_detections = np.concatenate([
image_boxes,
np.expand_dims(image_scores, axis=1),
np.expand_dims(image_labels, axis=1)
],
axis=1)
#drape boxes
#get image name
pc = postprocessing.drape_boxes(
boxes=quality_boxes,
tilename=generator.image_data[i]["tile"],
lidar_dir=DeepForest_config["lidar_path"])
#Skip if point density is too low
if pc:
#Get new bounding boxes
new_boxes = postprocessing.cloud_to_box(pc)
if save_path is not None:
draw_annotations(raw_image,
generator.load_annotations(i),
label_to_name=generator.label_to_name)
draw_detections(raw_image,
image_boxes,
image_scores,
image_labels,
label_to_name=generator.label_to_name,
score_threshold=score_threshold)
image_name = generator.image_names[i]
row = generator.image_data[image_name]
fname = os.path.splitext(row["tile"])[0] + "_" + str(row["window"])
#Write RGB
cv2.imwrite(os.path.join(save_path, '{}.png'.format(fname)),
raw_image)
if experiment:
experiment.log_image(os.path.join(save_path,
'{}.png'.format(fname)),
file_name=fname)
# copy detections to all_detections
for label in range(generator.num_classes()):
all_detections[i][label] = image_detections[
image_detections[:, -1] == label, :-1]
#print('{}/{}'.format(i + 1, generator.size()), end='\r')
return all_detections
def _get_detections(generator, model, score_threshold=0.05, max_detections=100, save_path=None, experiment=None):
""" Get the detections from the model using the generator.
The result is a list of lists such that the size is:
all_detections[num_images][num_classes] = detections[num_detections, 4 + num_classes]
# Arguments
generator : The generator used to run images through the model.
model : The model to run on the images.
score_threshold : The score confidence threshold to use.
max_detections : The maximum number of detections to use per image.
save_path : The path to save the images with visualized detections to.
experiment : Comet ML experiment
# Returns
A list of lists containing the detections for each image in the generator.
"""
all_detections = [[None for i in range(generator.num_classes())] for j in range(generator.size())]
for i in range(generator.size()):
raw_image = generator.load_image(i)
plot_image = copy.deepcopy(raw_image)
#Format name and save
image_name = generator.image_names[i]
row = generator.image_data[image_name]
lfname = os.path.splitext(row["tile"])[0] + "_" + str(row["window"]) +"raw_image"
#Skip if missing a component data source
if raw_image is False:
print("Empty image, skipping")
continue
#Store plotting images
plot_rgb = plot_image[:,:,:3].copy()
plot_chm = plot_image[:,:,3]
#predict
image = generator.preprocess_image(raw_image)
image, scale = generator.resize_image(image)
if keras.backend.image_data_format() == 'channels_first':
image = image.transpose((2, 0, 1))
# run network
boxes, scores, labels = model.predict_on_batch(np.expand_dims(image, axis=0))[:3]
# correct boxes for image scale
boxes /= scale
# select indices which have a score above the threshold
indices = np.where(scores[0, :] > score_threshold)[0]
# select those scores
scores = scores[0][indices]
# find the order with which to sort the scores
scores_sort = np.argsort(-scores)[:max_detections]
# select detections
image_boxes = boxes[0, indices[scores_sort], :]
image_scores = scores[scores_sort]
image_labels = labels[0, indices[scores_sort]]
image_detections = np.concatenate([image_boxes, np.expand_dims(image_scores, axis=1), np.expand_dims(image_labels, axis=1)], axis=1)
#name image
image_name = generator.image_names[i]
row = generator.image_data[image_name]
fname = os.path.splitext(row["tile"])[0] + "_" + str(row["window"])
#drape boxes
#get lidar cloud if a new tile, or if not the same tile as previous image.
if generator.with_lidar:
if i == 0:
generator.load_lidar_tile()
elif not generator.image_data[i]["tile"] == generator.image_data[i-1]["tile"]:
generator.load_lidar_tile()
#The tile could be the full tile, so let's check just the 400 pixel crop we are interested
#Not the best structure, but the on-the-fly generator always has 0 bounds
if hasattr(generator, 'hf'):
bounds = generator.hf["utm_coords"][generator.row["window"]]
else:
bounds=[]
if generator.with_lidar:
density = Lidar.check_density(generator.lidar_tile, bounds=bounds)
if density > 100:
#find window utm coordinates
#print("Bounds for image {}, window {}, are {}".format(generator.row["tile"], generator.row["window"], bounds))
pc = postprocessing.drape_boxes(boxes=image_boxes, pc = generator.lidar_tile, bounds=bounds)
#Get new bounding boxes
image_boxes = postprocessing.cloud_to_box(pc, bounds)
image_scores = image_scores[:image_boxes.shape[0]]
image_labels = image_labels[:image_boxes.shape[0]]
image_detections = np.concatenate([image_boxes, np.expand_dims(image_scores, axis=1), np.expand_dims(image_labels, axis=1)], axis=1)
else:
pass
#print("Point density of {:.2f} is too low, skipping image {}".format(density, generator.row["tile"]))
if save_path is not None:
draw_annotations(plot_rgb, generator.load_annotations(i), label_to_name=generator.label_to_name)
draw_detections(plot_rgb, image_boxes, image_scores, image_labels, label_to_name=generator.label_to_name,score_threshold=score_threshold)
#name image
image_name=generator.image_names[i]
row=generator.image_data[image_name]
fname=os.path.splitext(row["tile"])[0] + "_" + str(row["window"])
#Write RGB
cv2.imwrite(os.path.join(save_path, '{}.png'.format(fname)), plot_rgb)
#Format name and save
image_name = generator.image_names[i]
row = generator.image_data[image_name]
lfname = os.path.splitext(row["tile"])[0] + "_" + str(row["window"]) +"_lidar"
#make cv2 colormap
#normalize visual to make clearer for plotting
plot_chm = plot_chm/plot_chm.max() * 255
chm = np.uint8(plot_chm)
draw_annotations(chm, generator.load_annotations(i), label_to_name=generator.label_to_name)
draw_detections(chm, image_boxes, image_scores, image_labels, label_to_name=generator.label_to_name,score_threshold=score_threshold)
#Write CHM
cv2.imwrite(os.path.join(save_path, '{}_LIDAR.png'.format(lfname)), chm)
if experiment:
experiment.log_image(os.path.join(save_path, '{}_LIDAR.png'.format(lfname)),file_name=lfname)
experiment.log_image(os.path.join(save_path, '{}.png'.format(fname)),file_name=fname)
# copy detections to all_detections
for label in range(generator.num_classes()):
all_detections[i][label] = image_detections[image_detections[:, -1] == label, :-1]
return all_detections
# scores are sorted so we can break
if score < args.score_threshold:
break
#drape boxes
#get image name and load point cloud
image_name = os.path.splitext(os.path.basename(image_path))[0]
point_cloud_filename = os.path.join(DeepForest_config["lidar_path"] +
image_name) + ".laz"
pc = Lidar.load_lidar(point_cloud_filename)
pc = postprocessing.drape_boxes(boxes=quality_boxes, pc=pc)
#Skip if point density is too low
if pc:
#Get new bounding boxes
new_boxes = postprocessing.cloud_to_box(pc)
#expends 3dim
new_boxes = np.expand_dims(new_boxes, 0)
# visualize detections
for box, score, label in zip(new_boxes[0], scores[0], labels[0]):
# scores are sorted so we can break
if score < args.score_threshold:
break
color = label_color(label)
b = box.astype(int)
draw_box(draw, b, color=(0, 90, 255))
#caption = "{} {:.2f}".format(labels_to_names[label], score)
def neonRecall(
sites,
generator,
model,
score_threshold=0.05,
max_detections=100,
suppression_threshold=0.15,
save_path=None,
experiment=None):
point_contains = [ ]
site_data_dict = {}
for site in sites:
#Container for recall pts.
#load field data
field_data = pd.read_csv("data/field_data.csv")
field_data = field_data[field_data['UTM_E'].notnull()]
#select site
site_data = field_data[field_data["siteID"]==site]
#select tree species
specieslist = pd.read_csv("data/AcceptedSpecies.csv",encoding="utf-8")
specieslist = specieslist[specieslist["siteID"] == site]
site_data = site_data[site_data["scientificName"].isin(specieslist["scientificName"].values)]
#Single bole individuals as representitve, no individualID ending in non-digits
site_data = site_data[site_data["individualID"].str.contains("\d$")]
site_data_dict[site] = site_data
#Only data within the last two years, sites can be hand managed
#site_data=site_data[site_data["eventID"].str.contains("2015|2016|2017|2018")]
for i in range(generator.size()):
#Load image
raw_image = generator.load_image(i)
plot_image = copy.deepcopy(raw_image)
#Skip if missing a component data source
if raw_image is False:
print("Empty image, skipping")
continue
#Store plotting images.
plot_rgb = plot_image[:,:,:3].copy()
plot_chm = plot_image[:,:,3]
image = generator.preprocess_image(raw_image)
image, scale = generator.resize_image(image)
# run network
boxes, scores, labels = model.predict_on_batch(np.expand_dims(image, axis=0))[:3]
# correct boxes for image scale
boxes /= scale
# select indices which have a score above the threshold
indices = np.where(scores[0, :] > score_threshold)[0]
# select those scores
scores = scores[0][indices]
# find the order with which to sort the scores
scores_sort = np.argsort(-scores)[:max_detections]
# select detections
image_boxes = boxes[0, indices[scores_sort], :]
image_scores = scores[scores_sort]
image_labels = labels[0, indices[scores_sort]]
image_detections = np.concatenate([image_boxes, np.expand_dims(image_scores, axis=1), np.expand_dims(image_labels, axis=1)], axis=1)
#Find geographic bounds
base_dir = generator.DeepForest_config[generator.row["site"]][generator.name]["RGB"]
tile_path = os.path.join(base_dir, generator.image_data[i]["tile"])
with rasterio.open(tile_path) as dataset:
tile_bounds = dataset.bounds
#drape boxes
#get lidar cloud if a new tile, or if not the same tile as previous image.
if i == 0:
generator.load_lidar_tile()
elif not generator.image_data[i]["tile"] == generator.image_data[i-1]["tile"]:
generator.load_lidar_tile()
#The tile could be the full tile, so let's check just the 400 pixel crop we are interested
#Not the best structure, but the on-the-fly generator always has 0 bounds
if hasattr(generator, 'hf'):
bounds = generator.hf["utm_coords"][generator.row["window"]]
else:
bounds=[]
density = Lidar.check_density(generator.lidar_tile, bounds=bounds)
#print("Point density is {:.2f}".format(density))
if density > 100:
#find window utm coordinates
#print("Bounds for image {}, window {}, are {}".format(generator.row["tile"], generator.row["window"], bounds))
pc = postprocessing.drape_boxes(boxes=image_boxes, pc = generator.lidar_tile, bounds=bounds)
#Get new bounding boxes
new_boxes = postprocessing.cloud_to_box(pc, bounds)
new_scores = image_scores[:new_boxes.shape[0]]
new_labels = image_labels[:new_boxes.shape[0]]
image_detections = np.concatenate([new_boxes, np.expand_dims(new_scores, axis=1), np.expand_dims(new_labels, axis=1)], axis=1)
else:
#print("Point density of {:.2f} is too low, skipping image {}".format(density, generator.row["tile"]))
pass
#add spatial NEON points
site_data =site_data_dict[generator.row["site"]]
plotID = os.path.splitext(generator.image_data[i]["tile"])[0]
plot_data = site_data[site_data.plotID == plotID]
#Save image and send it to logger
if save_path is not None:
x = (plot_data.UTM_E - tile_bounds.left).values / 0.1
y = (tile_bounds.top - plot_data.UTM_N).values / 0.1
for j in np.arange(len(x)):
cv2.circle(plot_image,(int(x[j]),int(y[j])), 2, (0,0,255), -1)
#Write RGB
draw_detections(plot_rgb, image_boxes, image_scores, image_labels, label_to_name=generator.label_to_name,score_threshold=score_threshold)
#name image
image_name=generator.image_names[i]
row=generator.image_data[image_name]
fname=os.path.splitext(row["tile"])[0] + "_" + str(row["window"])
#Write RGB
cv2.imwrite(os.path.join(save_path, '{}_NeonPlot.png'.format(fname)), plot_rgb)
plot_chm = plot_chm/plot_chm.max() * 255
chm = np.uint8(plot_chm)
draw_detections(chm, image_boxes, image_scores, image_labels, label_to_name=generator.label_to_name, score_threshold=score_threshold, color = (80,127,255))
cv2.imwrite(os.path.join(save_path, '{}_Lidar_NeonPlot.png'.format(plotID)), chm)
#Format name and save
if experiment:
experiment.log_image(os.path.join(save_path, '{}_NeonPlot.png'.format(plotID)),file_name=str(plotID))
experiment.log_image(os.path.join(save_path, '{}_Lidar_NeonPlot.png'.format(plotID)),file_name=str("Lidar_" + plotID))
#calculate recall
s = gp.GeoSeries(map(Point, zip(plot_data.UTM_E, plot_data.UTM_N)))
#Calculate recall
projected_boxes = []
for row in image_boxes:
#Add utm bounds and create a shapely polygon
pbox=create_polygon(row, tile_bounds, cell_size=0.1)
projected_boxes.append(pbox)
#for each point, is it within a prediction?
for index, tree in plot_data.iterrows():
p=Point(tree.UTM_E, tree.UTM_N)
within_polygon=[]
for prediction in projected_boxes:
within_polygon.append(p.within(prediction))
#Check for overlapping polygon, add it to list
is_within = sum(within_polygon) > 0
point_contains.append(is_within)
#sum recall across plots
if len(point_contains)==0:
recall = None
else:
## Recall rate for plot
recall = sum(point_contains)/len(point_contains)
return(recall)