Skip to content

Support multiple neural networks at once #1040

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
m3d merged 1 commit into from
Feb 27, 2026
Merged

Support multiple neural networks at once #1040

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
88 changes: 56 additions & 32 deletions osgar/drivers/oak_camera_v3.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -66,14 +66,16 @@ def __init__(self, config, bus):
self.sleep_on_start_sec = config.get('sleep_on_start_sec')
self.verbose_detections = config.get('verbose_detections', True)

# NN setup variables
self.oak_config_model = config.get("model")
self.oak_config_nn_config = config.get("nn_config", {})
self.oak_config_nn_family = self.oak_config_nn_config.get("NN_family", "YOLO")
self.oak_config_nn_image_size = None
if "input_size" in self.oak_config_nn_config:
self.oak_config_nn_image_size = tuple(map(int, self.oak_config_nn_config.get("input_size").split('x')))
self.labels = config.get("mappings", {}).get("labels", [])
# NN setup variables - support for multiple models
self.models = config.get("models", [])

# Backwards compatibility for single-model configs
if "model" in config and len(self.models) == 0:
self.models.append({
"model": config.get("model"),
"nn_config": config.get("nn_config", {}),
"mappings": config.get("mappings", {})
})

def start(self):
if self.sleep_on_start_sec is not None:
Expand Down Expand Up @@ -165,20 +167,30 @@ def process(self, frame):
odom.transform.link(slam.odom)
gridmap_queue = slam.occupancyGridMap.createOutputQueue(blocking=False)

# Configure Neural Network processing block
if self.oak_config_model is not None:
nnPath = self.oak_config_model.get("blob")
assert Path(nnPath).exists(), "No blob found at '{}'!".format(nnPath)
W, H = self.oak_config_nn_image_size
# Configure Neural Network processing blocks dynamically
nn_queues = []
for model_cfg in self.models:
nn_path = model_cfg.get("model", {}).get("blob")
if not nn_path:
continue
assert Path(nn_path).exists(), "No blob found at '{}'!".format(nn_path)

nn_config = model_cfg.get("nn_config", {})
nn_family = nn_config.get("NN_family", "YOLO")

input_size_str = nn_config.get("input_size")
if input_size_str:
W, H = tuple(map(int, input_size_str.split('x')))
else:
W, H = 416, 416 # default fallback

nn = pipeline.create(dai.node.NeuralNetwork)
nn.setBlobPath(nnPath)
nn.setBlobPath(nn_path)
nn.setNumInferenceThreads(2)
nn.input.setBlocking(False)

# In DepthAI v3, YoloDetectionNetwork is replaced by NeuralNetwork + DetectionParser
if self.oak_config_nn_family == 'YOLO':
metadata = self.oak_config_nn_config.get("NN_specific_metadata", {})
if nn_family == 'YOLO':
metadata = nn_config.get("NN_specific_metadata", {})
parser = pipeline.create(dai.node.DetectionParser)
if "confidence_threshold" in metadata:
parser.setConfidenceThreshold(metadata["confidence_threshold"])
Expand All @@ -194,63 +206,75 @@ def process(self, frame):
parser.setIouThreshold(metadata["iou_threshold"])

nn.out.link(parser.input)
nn_queue = parser.out.createOutputQueue(blocking=False)
queue = parser.out.createOutputQueue(blocking=False)
else:
nn_queue = nn.out.createOutputQueue(blocking=False)
queue = nn.out.createOutputQueue(blocking=False)

# Feed from color camera via specific requested output resolution
# Feed from color camera via specific requested output resolution per model
nn_cam_out = cam_rgb.requestOutput((W, H), type=dai.ImgFrame.Type.BGR888p)
nn_cam_out.link(nn.input)

nn_queues.append({
"queue": queue,
"family": nn_family,
"image_size": (W, H),
"labels": model_cfg.get("mappings", {}).get("labels", [])
})


# Connect to device and start pipeline
pipeline.start()

while pipeline.isRunning() and self.bus.is_alive():
processed_any = False

# 1. Check Neural Networks
if self.oak_config_model is not None:
nn_packets = nn_queue.tryGetAll()
for nn_setup in nn_queues:
queue = nn_setup["queue"]
nn_family = nn_setup["family"]
W, H = nn_setup["image_size"]
labels = nn_setup["labels"]

nn_packets = queue.tryGetAll()
if nn_packets and len(nn_packets) > 0:
processed_any = True
packet = nn_packets[-1] # use latest packet
seq_num = packet.getSequenceNum()
dt = packet.getTimestamp()
timestamp_us = ((dt.days * 24 * 3600 + dt.seconds) * 1000000 + dt.microseconds)

if self.oak_config_nn_family == 'YOLO':
if nn_family == 'YOLO':
detections = packet.detections
bbox_list = []
for detection in detections:
bbox = (detection.xmin, detection.ymin, detection.xmax, detection.ymax)
if self.verbose_detections:
print(self.labels[detection.label], detection.confidence, bbox)
bbox_list.append([self.labels[detection.label], detection.confidence, list(bbox)])
print(labels[detection.label], detection.confidence, bbox)
bbox_list.append([labels[detection.label], detection.confidence, list(bbox)])
self.bus.publish("detections_seq", [seq_num, timestamp_us])
self.bus.publish('detections', bbox_list)

elif self.oak_config_nn_family == 'resnet':
elif nn_family == 'resnet':
nn_output = packet.getLayerFp16('output')
WIDTH, HEIGHT = self.oak_config_nn_image_size
mask = np.array(nn_output).reshape((2, HEIGHT, WIDTH))
mask = np.array(nn_output).reshape((2, H, W))
mask = mask.argmax(0).astype(np.uint8)
self.bus.publish('nn_mask', mask)

elif self.oak_config_nn_family == 'robotourist':
elif nn_family == 'robotourist' or nn_family == 'redroad':
# v3: getTensor automatically returns a NumPy array
nn_output = packet.getTensor('redroad_output')
nn_robotourist_output = packet.getTensor('robotourist_output')
WIDTH, HEIGHT = self.oak_config_nn_image_size

redroad = np.array(nn_output).reshape((HEIGHT//2, WIDTH//2))
redroad = np.array(nn_output).reshape((H//2, W//2))
self.bus.publish('redroad', redroad)
mask = (redroad > 0).astype(np.uint8)
mask = np.array(mask).reshape((HEIGHT//2, WIDTH//2))
mask = np.array(mask).reshape((H//2, W//2))
self.bus.publish('nn_mask', mask)

robotourist = np.array(nn_robotourist_output).reshape((1280, 7, 7))
self.bus.publish('robotourist', robotourist)


# 2. Check Depth
if self.is_depth:
depth_frames = depth_queue.tryGetAll()
Expand Down