azaion/annotations
1
0
Fork 0
mirror of https://github.com/azaion/annotations.git synced 2026-04-22 10:46:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

fix loader bug with _CACHED_HW_INFO

Browse Source 
put tile size to name and set it dynamically for AI recognition
This commit is contained in:
Oleksandr Bezdieniezhnykh
2025-09-02 13:59:23 +03:00
parent 067f02cc63
commit b6b6751c37
19 changed files with 83 additions and 104 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Azaion.Inference/inference.pyx
+18 -23
View File
@@ -58,8 +58,6 @@ cdef class Inference:
self.model_input = None
self.model_width = 0
self.model_height = 0
self.tile_width = 0
self.tile_height = 0
self.engine = None
self.is_building_engine = False
self.ai_availability_status = AIAvailabilityStatus()
@@ -107,15 +105,11 @@ cdef class Inference:
self.is_building_engine = False
self.model_height, self.model_width = self.engine.get_input_shape()
#todo: temporarily, send it from the client
self.tile_width = 550
self.tile_height = 550
except Exception as e:
self.ai_availability_status.set_status(AIAvailabilityEnum.ERROR, <str>str(e))
self.is_building_engine = False
cdef preprocess(self, frames):
blobs = [cv2.dnn.blobFromImage(frame,
scalefactor=1.0 / 255.0,
@@ -277,7 +271,7 @@ cdef class Inference:
if img_h <= 1.5 * self.model_height and img_w <= 1.5 * self.model_width:
frame_data.append((frame, original_media_name, f'{original_media_name}_000000'))
else:
res = self.split_to_tiles(frame, path, ai_config.big_image_tile_overlap_percent)
res = self.split_to_tiles(frame, path, ai_config.tile_size, ai_config.big_image_tile_overlap_percent)
frame_data.extend(res)
if len(frame_data) > self.engine.get_batch_size():
for chunk in self.split_list_extend(frame_data, self.engine.get_batch_size()):
@@ -287,31 +281,31 @@ cdef class Inference:
self._process_images_inner(cmd, ai_config, chunk)
cdef split_to_tiles(self, frame, path, overlap_percent):
cdef split_to_tiles(self, frame, path, tile_size, overlap_percent):
constants_inf.log(<str>f'splitting image {path} to tiles...')
img_h, img_w, _ = frame.shape
stride_w = int(self.tile_width * (1 - overlap_percent / 100))
stride_h = int(self.tile_height * (1 - overlap_percent / 100))
stride_w = int(tile_size * (1 - overlap_percent / 100))
stride_h = int(tile_size * (1 - overlap_percent / 100))
results = []
original_media_name = Path(<str> path).stem.replace(" ", "")
for y in range(0, img_h, stride_h):
for x in range(0, img_w, stride_w):
x_end = min(x + self.tile_width, img_w)
y_end = min(y + self.tile_height, img_h)
x_end = min(x + tile_size, img_w)
y_end = min(y + tile_size, img_h)
# correct x,y for the close-to-border tiles
if x_end - x < self.tile_width:
if img_w - (x - stride_w) <= self.tile_width:
if x_end - x < tile_size:
if img_w - (x - stride_w) <= tile_size:
continue # the previous tile already covered the last gap
x = img_w - self.tile_width
if y_end - y < self.tile_height:
if img_h - (y - stride_h) <= self.tile_height:
x = img_w - tile_size
if y_end - y < tile_size:
if img_h - (y - stride_h) <= tile_size:
continue # the previous tile already covered the last gap
y = img_h - self.tile_height
y = img_h - tile_size
tile = frame[y:y_end, x:x_end]
name = f'{original_media_name}{constants_inf.SPLIT_SUFFIX}{x:04d}_{y:04d}!_000000'
name = f'{original_media_name}{constants_inf.SPLIT_SUFFIX}{tile_size:04d}{x:04d}_{y:04d}!_000000'
results.append((tile, original_media_name, name))
return results
@@ -337,14 +331,15 @@ cdef class Inference:
cdef remove_tiled_duplicates(self, Annotation annotation):
right = annotation.name.rindex('!')
left = annotation.name.index(constants_inf.SPLIT_SUFFIX) + len(constants_inf.SPLIT_SUFFIX)
x_str, y_str = annotation.name[left:right].split('_')
tile_size_str, x_str, y_str = annotation.name[left:right].split('_')
tile_size = int(tile_size_str)
x = int(x_str)
y = int(y_str)
for det in annotation.detections:
x1 = det.x * self.tile_width
y1 = det.y * self.tile_height
det_abs = Detection(x + x1, y + y1, det.w * self.tile_width, det.h * self.tile_height, det.cls, det.confidence)
x1 = det.x * tile_size
y1 = det.y * tile_size
det_abs = Detection(x + x1, y + y1, det.w * tile_size, det.h * tile_size, det.cls, det.confidence)
detections = self._tile_detections.setdefault(annotation.original_media_name, [])
if det_abs in detections:
annotation.detections.remove(det)