YOLO Node Stops Executing

LorenzoR

When running a stereo pipeline that uses a YOLO detection node along with depth estimation and feature tracking, the YOLO branch initially processes frames (≈11 FPS) but then stops executing after a few seconds.

At that point, the depth branch reaches 30 FPS and the NOC DDR bandwidth drops (from ~1300 MB/s to ~430 MB/s). This suggests that the YOLO node is no longer processing frames.

I tried setting a different number of inference threads, recompiling with a different number of shaves, etc, but nothing seems to solve it.

Script to reproduce the faulty behavior:

import depthai
import time
import cv2

MODEL_NAME = "utils/dai-yolo/yolov8n_coco_640x352.blob"

nnPath = MODEL_NAME

class DepthAIProcessor:
    def __init__(self):
        pipeline = depthai.Pipeline()

        # YOLO setup
        detectionNetwork = pipeline.create(depthai.node.YoloDetectionNetwork)

        detectionNetwork.setConfidenceThreshold(0.5)
        detectionNetwork.setNumClasses(80)
        detectionNetwork.setCoordinateSize(4)
        detectionNetwork.setIouThreshold(0.5)
        detectionNetwork.setBlobPath(nnPath)
        detectionNetwork.setNumInferenceThreads(2)
        detectionNetwork.input.setQueueSize(1)
        detectionNetwork.input.setBlocking(False)

        nnOut = pipeline.create(depthai.node.XLinkOut)
        nnOut.setStreamName("nn")

        manipLeft = pipeline.create(depthai.node.ImageManip)
        manipLeft.inputImage.setBlocking(False)
        manipLeft.inputImage.setQueueSize(1)
        #manipLeft.initialConfig.setResizeThumbnail(416, 416)
        manipLeft.initialConfig.setResizeThumbnail(640, 352)
        manipLeft.initialConfig.setFrameType(depthai.RawImgFrame.Type.RGB888p)

        manipLeft.out.link(detectionNetwork.input)
        detectionNetwork.out.link(nnOut.input)

        # Depth output
        xoutDepth = pipeline.create(depthai.node.XLinkOut)
        xoutDepth.setStreamName("depth")
        xoutDepth.input.setBlocking(False)
        xoutDepth.input.setQueueSize(3)

        xoutManip = pipeline.create(depthai.node.XLinkOut)
        xoutManip.setStreamName("manip")
        xoutManip.input.setQueueSize(1)
        xoutManip.input.setBlocking(False)
        manipLeft.out.link(xoutManip.input)

        stereo = pipeline.create(depthai.node.StereoDepth)
        stereo.setDepthAlign(depthai.StereoDepthProperties.DepthAlign.RECTIFIED_LEFT)
        stereo.depth.link(xoutDepth.input)   

        FPS = 30
        monoLeft = pipeline.create(depthai.node.MonoCamera)
        monoLeft.setCamera("left")
        monoLeft.setFps(FPS)
        monoLeft.setResolution(depthai.MonoCameraProperties.SensorResolution.THE_400_P)

        monoLeft.out.link(stereo.left)

        monoRight = pipeline.create(depthai.node.MonoCamera)
        monoRight.setCamera("right")
        monoRight.setFps(FPS)
        monoRight.setResolution(depthai.MonoCameraProperties.SensorResolution.THE_400_P)
        monoRight.out.link(stereo.right)


        featureTrackerLeft = pipeline.create(depthai.node.FeatureTracker)
        monoLeft.out.link(featureTrackerLeft.inputImage)
        xoutFeaturesLeft = pipeline.create(depthai.node.XLinkOut)
        featureTrackerLeft.outputFeatures.link(xoutFeaturesLeft.input) # left feature tracker
        xoutFeaturesLeft.setStreamName("trackedFeaturesLeft")


        featureTrackerRight = pipeline.create(depthai.node.FeatureTracker)
        monoRight.out.link(featureTrackerRight.inputImage)
        xoutFeaturesRight = pipeline.create(depthai.node.XLinkOut)
        featureTrackerRight.outputFeatures.link(xoutFeaturesRight.input) # right feature tracker
        xoutFeaturesRight.setStreamName("trackedFeaturesRight")

        stereo.rectifiedLeft.link(manipLeft.inputImage)

        nnOut.input.setBlocking(False)
        nnOut.input.setQueueSize(1)

        self.device = depthai.Device(pipeline)

        self.nn_queue = self.device.getOutputQueue(name="nn", maxSize=1, blocking=False)
        self.depth_queue = self.device.getOutputQueue(name="depth", maxSize=3, blocking=False)
        self.left_features_queue = self.device.getOutputQueue(name="trackedFeaturesLeft", maxSize=1, blocking=False)
        self.right_features_queue = self.device.getOutputQueue(name="trackedFeaturesRight", maxSize=1, blocking=False)
        self.manip_queue = self.device.getOutputQueue(name="manip", maxSize=1, blocking=False)


    def run(self):
        depth_frame_count = 0
        last_fps_time = time.time()
        yolo_frame_count = 0
        yolo_last_fps_time = time.time()
        manip_frame_count = 0
        manip_last_fps_time = time.time()
        fps_interval = 1

        latest_features_left = 0
        latest_features_right = 0

        while True:
            try:                
                in_nn = self.nn_queue.tryGet()
                in_depth = self.depth_queue.tryGet()
                in_manip = self.manip_queue.tryGet()

                in_features_left = self.left_features_queue.tryGet()
                in_features_right = self.right_features_queue.tryGet()

                if in_features_left:
                    latest_features_left = len(in_features_left.trackedFeatures)

                if in_features_right:
                    latest_features_right = len(in_features_right.trackedFeatures)

                if in_depth is not None:
                    current_time = time.time()
                    data_depth = in_depth.getCvFrame()
                    cv2.imshow("depth", data_depth)


                    depth_frame_count += 1
                    if current_time - last_fps_time >= fps_interval:
                        fps = depth_frame_count / (current_time - last_fps_time)
                        print(f"Depth Map FPS: {fps:.2f}")
                        depth_frame_count = 0
                        last_fps_time = current_time

                        print(f"featuresRight {latest_features_right}")
                        print(f"featuresLeft {latest_features_left}")

                if in_nn is not None:
                    yolo_frame_count += 1
                    current_time = time.time()
                    if current_time - yolo_last_fps_time >= fps_interval:
                        yolo_fps = yolo_frame_count / (current_time - yolo_last_fps_time)
                        print(f"YOLO FPS: {yolo_fps:.2f}")
                        yolo_frame_count = 0
                        yolo_last_fps_time = current_time

                if in_manip is not None:
                    cv2.imshow("manip", in_manip.getCvFrame())
                    manip_frame_count += 1
                    current_time = time.time()
                    if current_time - manip_last_fps_time >= fps_interval:
                        manip_fps = manip_frame_count / (current_time - manip_last_fps_time)
                        print(f"Manip FPS: {manip_fps:.2f}")
                        manip_frame_count = 0
                        manip_last_fps_time = current_time

                cv2.waitKey(1)


            except KeyboardInterrupt:
                print("Stopping...")
                break
            except Exception as e:
                print(f"Error: {e}")
                continue

    def close(self):
        self.device.close()

def main():
    processor = DepthAIProcessor()
try:
    processor.run()
except KeyboardInterrupt:
    print("Keyboard Interrupt received")
finally:
    processor.close()

if __name__ == '__main__':
    main()

Example output:

$ DEPTHAI_LEVEL=info python3 script.py
[19443010319B191300] [1.1] [0.898] [system] [info] Memory Usage - DDR: 0.12 / 333.28 MiB, CMX: 2.04 / 2.50 MiB, LeonOS Heap: 7.51 / 81.76 MiB, LeonRT Heap: 2.89 / 39.90 MiB / NOC ddr: 26 MB/s
[19443010319B191300] [1.1] [0.898] [system] [info] Temperatures - Average: 43.73C, CSS: 45.77C, MSS 43.27C, UPA: 43.04C, DSS: 42.81C
[19443010319B191300] [1.1] [0.898] [system] [info] Cpu Usage - LeonOS 70.73%, LeonRT: 1.79%
[2025-02-14 15:17:54.425] [depthai] [info] Logging disabled
[19443010319B191300] [1.1] [1.017] [MonoCamera(6)] [info] Using board socket: -1, id: 1
[19443010319B191300] [1.1] [1.018] [MonoCamera(7)] [info] Using board socket: -1, id: 2
[19443010319B191300] [1.1] [1.018] [FeatureTracker(8)] [info] Init node feature tracker.
[19443010319B191300] [1.1] [1.018] [FeatureTracker(10)] [info] Init node feature tracker.
[19443010319B191300] [1.1] [1.021] [system] [info] SIPP (Signal Image Processing Pipeline) internal buffer size '18432'B, DMA buffer size: '16384'B
[19443010319B191300] [1.1] [1.056] [system] [info] ImageManip internal buffer size '363520'B, shave buffer size '35840'B
[19443010319B191300] [1.1] [1.056] [system] [info] NeuralNetwork allocated resources: shaves: [0-10] cmx slices: [0-10]
MonoCamera allocated resources: no shaves; cmx slices: [13-15]
StereoDepth allocated resources: shaves: [12-12] cmx slices: [11-11]
ImageManip allocated resources: shaves: [15-15] no cmx slices.
FeatureTracker allocated resources: shaves: [13-13] cmx slices: [12-12]

[19443010319B191300] [1.1] [1.069] [DetectionNetwork(0)] [info] Needed resources: shaves: 6, ddr: 4956160
[19443010319B191300] [1.1] [1.069] [DetectionNetwork(0)] [warning] Network compiled for 6 shaves, maximum available 11, compiling for 5 shaves likely will yield in better performance
[19443010319B191300] [1.1] [1.083] [DetectionNetwork(0)] [info] Inference thread count: 2, number of shaves allocated per thread: 6, number of Neural Compute Engines (NCE) allocated per thread: 1
[19443010319B191300] [1.1] [1.246] [StereoDepth(5)] [info] Depth baseline: '7.5000005' cm, focal length: '401.15234'
[19443010319B191300] [1.1] [1.246] [StereoDepth(5)] [info] Depth horizontal FOV: '77.15886', vertical FOV: '52.998356'
[19443010319B191300] [1.1] [1.246] [StereoDepth(5)] [info] Depth intrinsics:
401\.152,    0.000,  318.788,
0\.000,  401.152,  190.277,
0\.000,    0.000,    1.000,
[19443010319B191300] [1.1] [1.899] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 474 MB/s
[19443010319B191300] [1.1] [1.899] [system] [info] Temperatures - Average: 47.07C, CSS: 47.80C, MSS 46.68C, UPA: 48.03C, DSS: 45.77C
[19443010319B191300] [1.1] [1.899] [system] [info] Cpu Usage - LeonOS 46.05%, LeonRT: 44.70%
Depth Map FPS: 25.39
featuresRight 266
featuresLeft 264
Manip FPS: 22.36
YOLO FPS: 8.33
[19443010319B191300] [1.1] [2.900] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1343 MB/s
[19443010319B191300] [1.1] [2.900] [system] [info] Temperatures - Average: 47.97C, CSS: 48.03C, MSS 47.80C, UPA: 49.14C, DSS: 46.90C
[19443010319B191300] [1.1] [2.900] [system] [info] Cpu Usage - LeonOS 40.77%, LeonRT: 34.12%
Depth Map FPS: 21.85
featuresRight 283
featuresLeft 278
Manip FPS: 21.84
YOLO FPS: 12.06
[19443010319B191300] [1.1] [3.901] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1316 MB/s
[19443010319B191300] [1.1] [3.901] [system] [info] Temperatures - Average: 49.14C, CSS: 49.37C, MSS 48.47C, UPA: 50.25C, DSS: 48.47C
[19443010319B191300] [1.1] [3.901] [system] [info] Cpu Usage - LeonOS 41.16%, LeonRT: 34.04%
Depth Map FPS: 21.30
featuresRight 293
featuresLeft 274
Manip FPS: 21.33
YOLO FPS: 9.25
[19443010319B191300] [1.1] [4.902] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1320 MB/s
[19443010319B191300] [1.1] [4.902] [system] [info] Temperatures - Average: 49.53C, CSS: 50.03C, MSS 48.70C, UPA: 50.70C, DSS: 48.70C
[19443010319B191300] [1.1] [4.902] [system] [info] Cpu Usage - LeonOS 42.80%, LeonRT: 34.63%
Depth Map FPS: 22.81
featuresRight 290
featuresLeft 282
Manip FPS: 22.59
YOLO FPS: 10.99
[19443010319B191300] [1.1] [5.903] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1323 MB/s
[19443010319B191300] [1.1] [5.903] [system] [info] Temperatures - Average: 49.75C, CSS: 50.70C, MSS 49.14C, UPA: 50.70C, DSS: 48.47C
[19443010319B191300] [1.1] [5.903] [system] [info] Cpu Usage - LeonOS 41.98%, LeonRT: 34.75%
Depth Map FPS: 21.78
featuresRight 293
featuresLeft 282
Manip FPS: 21.95
YOLO FPS: 11.13

...

[19443010319B191300] [1.1] [43.942] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1327 MB/s
[19443010319B191300] [1.1] [43.942] [system] [info] Temperatures - Average: 53.71C, CSS: 53.99C, MSS 53.77C, UPA: 54.86C, DSS: 52.24C
[19443010319B191300] [1.1] [43.942] [system] [info] Cpu Usage - LeonOS 43.23%, LeonRT: 34.93%
Manip FPS: 22.47
Depth Map FPS: 22.37
featuresRight 295
featuresLeft 284
[19443010319B191300] [1.1] [44.943] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 1140 MB/s
[19443010319B191300] [1.1] [44.943] [system] [info] Temperatures - Average: 52.35C, CSS: 52.68C, MSS 52.46C, UPA: 53.55C, DSS: 50.70C
[19443010319B191300] [1.1] [44.943] [system] [info] Cpu Usage - LeonOS 43.76%, LeonRT: 30.77%
Manip FPS: 28.61
Depth Map FPS: 29.89
featuresRight 296
featuresLeft 284
[19443010319B191300] [1.1] [45.944] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.99 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 432 MB/s
[19443010319B191300] [1.1] [45.944] [system] [info] Temperatures - Average: 52.18C, CSS: 53.12C, MSS 51.80C, UPA: 53.33C, DSS: 50.47C
[19443010319B191300] [1.1] [45.944] [system] [info] Cpu Usage - LeonOS 44.08%, LeonRT: 28.20%
Manip FPS: 30.84
Depth Map FPS: 30.81
featuresRight 294
featuresLeft 281
[19443010319B191300] [1.1] [46.945] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 431 MB/s
[19443010319B191300] [1.1] [46.945] [system] [info] Temperatures - Average: 52.40C, CSS: 53.55C, MSS 52.02C, UPA: 53.33C, DSS: 50.70C
[19443010319B191300] [1.1] [46.945] [system] [info] Cpu Usage - LeonOS 42.78%, LeonRT: 27.95%
Manip FPS: 30.41
Depth Map FPS: 30.14
featuresRight 294
featuresLeft 282
[19443010319B191300] [1.1] [47.946] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 425 MB/s
[19443010319B191300] [1.1] [47.946] [system] [info] Temperatures - Average: 51.91C, CSS: 52.90C, MSS 51.58C, UPA: 52.90C, DSS: 50.25C
[19443010319B191300] [1.1] [47.946] [system] [info] Cpu Usage - LeonOS 41.35%, LeonRT: 27.99%
Manip FPS: 29.97
Depth Map FPS: 29.98
featuresRight 292
featuresLeft 286
[19443010319B191300] [1.1] [48.947] [system] [info] Memory Usage - DDR: 34.82 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 5.76 / 39.90 MiB / NOC ddr: 425 MB/s
[19443010319B191300] [1.1] [48.947] [system] [info] Temperatures - Average: 51.85C, CSS: 53.12C, MSS 51.14C, UPA: 52.68C, DSS: 50.47C
[19443010319B191300] [1.1] [48.947] [system] [info] Cpu Usage - LeonOS 41.84%, LeonRT: 27.74%

Thank you for your support

jakaskerl

LorenzoR
Sounds like the device runs out of frames in the pool because the inference takes too long. Try manually lowering the FPS to 3 and checking if it still freezes. I hope this lower FPS gives the nodes enough time to process the frames.

Thanks,
Jaka

LorenzoR

Hello Jaka, thank you for your reply.

in the real script (this is actually a stripped minimal example), I noticed I couldn't keep inference running at that rate so I implemented a script node to lower YOLO FPS while keeping left frames at 30 FPS (I need the depth map to be output at 30FPS).

YOLO node still crashes after a while

import depthai
import time
import cv2

MODEL_NAME = "utils/dai-yolo/yolov8n_coco_640x352.blob"

nnPath = MODEL_NAME

class DepthAIProcessor:
    def __init__(self):
        pipeline = depthai.Pipeline()

        # YOLO setup
        detectionNetwork = pipeline.create(depthai.node.YoloDetectionNetwork)

        detectionNetwork.setConfidenceThreshold(0.5)
        detectionNetwork.setNumClasses(80)
        detectionNetwork.setCoordinateSize(4)
        detectionNetwork.setIouThreshold(0.5)
        detectionNetwork.setBlobPath(nnPath)
        detectionNetwork.setNumInferenceThreads(2)
        detectionNetwork.input.setQueueSize(1)
        detectionNetwork.input.setBlocking(False)

        nnOut = pipeline.create(depthai.node.XLinkOut)
        nnOut.setStreamName("nn")

        script = pipeline.create(depthai.node.Script)
        script.setScript("""
import time
FPS_YOLO = 3
                         
interval = 1.0 / FPS_YOLO
                               
last_send_time = Clock.now()
                         
while True:
    in_frame = node.io['in'].tryGet()
    if in_frame is not None:
        current_time = Clock.now()
        if (current_time - last_send_time).total_seconds() >= interval:
            last_send_time = current_time
            node.io['out'].send(in_frame)
    else:
        time.sleep(0.004) # need to sleep to avoid blocking everything!!!
""")


        script.inputs['in'].setBlocking(False)
        script.inputs['in'].setQueueSize(1)

        manipLeft = pipeline.create(depthai.node.ImageManip)
        manipLeft.inputImage.setBlocking(False)
        manipLeft.inputImage.setQueueSize(1)
        #manipLeft.initialConfig.setResizeThumbnail(416, 416)
        manipLeft.initialConfig.setResizeThumbnail(640, 352)
        manipLeft.initialConfig.setFrameType(depthai.RawImgFrame.Type.RGB888p)

        manipLeft.out.link(script.inputs['in'])
        #manipLeft.out.link(detectionNetwork.input)
        script.outputs['out'].link(detectionNetwork.input)
        detectionNetwork.out.link(nnOut.input)

        # Depth output
        xoutDepth = pipeline.create(depthai.node.XLinkOut)
        xoutDepth.setStreamName("depth")
        xoutDepth.input.setBlocking(False)
        xoutDepth.input.setQueueSize(3)

        xoutManip = pipeline.create(depthai.node.XLinkOut)
        xoutManip.setStreamName("manip")
        xoutManip.input.setQueueSize(1)
        xoutManip.input.setBlocking(False)
        manipLeft.out.link(xoutManip.input)

        stereo = pipeline.create(depthai.node.StereoDepth)
        stereo.setDepthAlign(depthai.StereoDepthProperties.DepthAlign.RECTIFIED_LEFT)
        stereo.depth.link(xoutDepth.input)   

        FPS = 30
        monoLeft = pipeline.create(depthai.node.MonoCamera)
        monoLeft.setCamera("left")
        monoLeft.setFps(FPS)
        monoLeft.setResolution(depthai.MonoCameraProperties.SensorResolution.THE_400_P)

        monoLeft.out.link(stereo.left)

        monoRight = pipeline.create(depthai.node.MonoCamera)
        monoRight.setCamera("right")
        monoRight.setFps(FPS)
        monoRight.setResolution(depthai.MonoCameraProperties.SensorResolution.THE_400_P)
        monoRight.out.link(stereo.right)


        featureTrackerLeft = pipeline.create(depthai.node.FeatureTracker)
        monoLeft.out.link(featureTrackerLeft.inputImage)
        xoutFeaturesLeft = pipeline.create(depthai.node.XLinkOut)
        featureTrackerLeft.outputFeatures.link(xoutFeaturesLeft.input) # left feature tracker
        xoutFeaturesLeft.setStreamName("trackedFeaturesLeft")


        featureTrackerRight = pipeline.create(depthai.node.FeatureTracker)
        monoRight.out.link(featureTrackerRight.inputImage)
        xoutFeaturesRight = pipeline.create(depthai.node.XLinkOut)
        featureTrackerRight.outputFeatures.link(xoutFeaturesRight.input) # right feature tracker
        xoutFeaturesRight.setStreamName("trackedFeaturesRight")


        stereo.rectifiedLeft.link(manipLeft.inputImage)

        nnOut.input.setBlocking(False)
        nnOut.input.setQueueSize(1)

        self.device = depthai.Device(pipeline)

        self.nn_queue = self.device.getOutputQueue(name="nn", maxSize=1, blocking=False)
        self.depth_queue = self.device.getOutputQueue(name="depth", maxSize=3, blocking=False)
        self.left_features_queue = self.device.getOutputQueue(name="trackedFeaturesLeft", maxSize=1, blocking=False)
        self.right_features_queue = self.device.getOutputQueue(name="trackedFeaturesRight", maxSize=1, blocking=False)
        self.manip_queue = self.device.getOutputQueue(name="manip", maxSize=1, blocking=False)


    def run(self):
        depth_frame_count = 0
        last_fps_time = time.time()
        yolo_frame_count = 0
        yolo_last_fps_time = time.time()
        manip_frame_count = 0
        manip_last_fps_time = time.time()
        fps_interval = 1

        latest_features_left = 0
        latest_features_right = 0

        while True:
            try:                
                in_nn = self.nn_queue.tryGet()
                in_depth = self.depth_queue.tryGet()
                in_manip = self.manip_queue.tryGet()

                in_features_left = self.left_features_queue.tryGet()
                in_features_right = self.right_features_queue.tryGet()

                if in_features_left:
                    latest_features_left = len(in_features_left.trackedFeatures)

                if in_features_right:
                    latest_features_right = len(in_features_right.trackedFeatures)


                if in_depth is not None:
                    current_time = time.time()
                    data_depth = in_depth.getCvFrame()
                    cv2.imshow("depth", data_depth)


                    depth_frame_count += 1
                    if current_time - last_fps_time >= fps_interval:
                        fps = depth_frame_count / (current_time - last_fps_time)
                        print(f"Depth Map FPS: {fps:.2f}")
                        depth_frame_count = 0
                        last_fps_time = current_time

                        print(f"featuresRight {latest_features_right}")
                        print(f"featuresLeft {latest_features_left}")

                if in_nn is not None:
                    yolo_frame_count += 1
                    current_time = time.time()
                    if current_time - yolo_last_fps_time >= fps_interval:
                        yolo_fps = yolo_frame_count / (current_time - yolo_last_fps_time)
                        print(f"YOLO FPS: {yolo_fps:.2f}")
                        yolo_frame_count = 0
                        yolo_last_fps_time = current_time

                if in_manip is not None:
                    cv2.imshow("manip", in_manip.getCvFrame())
                    manip_frame_count += 1
                    current_time = time.time()
                    if current_time - manip_last_fps_time >= fps_interval:
                        manip_fps = manip_frame_count / (current_time - manip_last_fps_time)
                        print(f"Manip FPS: {manip_fps:.2f}")
                        manip_frame_count = 0
                        manip_last_fps_time = current_time

                cv2.waitKey(1)


            except KeyboardInterrupt:
                print("Stopping...")
                break
            except Exception as e:
                print(f"Error: {e}")
                continue

    def close(self):
        self.device.close()

def main():
    processor = DepthAIProcessor()
    try:
        processor.run()
    except KeyboardInterrupt:
        print("Keyboard Interrupt received")
    finally:
        processor.close()

if __name__ == '__main__':
    main()

[19443010319B191300] [1.1] [44.956] [system] [info] Memory Usage - DDR: 34.83 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 9.19 / 39.90 MiB / NOC ddr: 674 MB/s
[19443010319B191300] [1.1] [44.956] [system] [info] Temperatures - Average: 44.87C, CSS: 45.32C, MSS 44.87C, UPA: 45.32C, DSS: 43.96C
[19443010319B191300] [1.1] [44.956] [system] [info] Cpu Usage - LeonOS 41.44%, LeonRT: 34.15%
Manip FPS: 30.00
Depth Map FPS: 29.95
featuresRight 264
featuresLeft 264
YOLO FPS: 2.91
[19443010319B191300] [1.1] [45.957] [system] [info] Memory Usage - DDR: 34.83 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.99 / 81.76 MiB, LeonRT Heap: 9.19 / 39.90 MiB / NOC ddr: 673 MB/s
[19443010319B191300] [1.1] [45.957] [system] [info] Temperatures - Average: 45.32C, CSS: 46.23C, MSS 45.09C, UPA: 45.77C, DSS: 44.18C
[19443010319B191300] [1.1] [45.957] [system] [info] Cpu Usage - LeonOS 41.63%, LeonRT: 34.32%
Manip FPS: 29.98
Depth Map FPS: 30.06
featuresRight 266
featuresLeft 262
[19443010319B191300] [1.1] [46.958] [system] [info] Memory Usage - DDR: 34.83 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 9.19 / 39.90 MiB / NOC ddr: 617 MB/s
[19443010319B191300] [1.1] [46.958] [system] [info] Temperatures - Average: 44.75C, CSS: 45.55C, MSS 44.18C, UPA: 45.32C, DSS: 43.96C
[19443010319B191300] [1.1] [46.958] [system] [info] Cpu Usage - LeonOS 41.80%, LeonRT: 32.14%
Manip FPS: 29.97
Depth Map FPS: 30.02
featuresRight 265
featuresLeft 263
[19443010319B191300] [1.1] [47.959] [system] [info] Memory Usage - DDR: 34.83 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 9.19 / 39.90 MiB / NOC ddr: 426 MB/s
[19443010319B191300] [1.1] [47.959] [system] [info] Temperatures - Average: 44.86C, CSS: 46.45C, MSS 44.41C, UPA: 45.32C, DSS: 43.27C
[19443010319B191300] [1.1] [47.959] [system] [info] Cpu Usage - LeonOS 41.43%, LeonRT: 30.89%
Manip FPS: 30.07
Depth Map FPS: 30.00
featuresRight 246
featuresLeft 266
[19443010319B191300] [1.1] [48.960] [system] [info] Memory Usage - DDR: 34.83 / 333.28 MiB, CMX: 2.47 / 2.50 MiB, LeonOS Heap: 35.97 / 81.76 MiB, LeonRT Heap: 9.19 / 39.90 MiB / NOC ddr: 424 MB/s
[19443010319B191300] [1.1] [48.960] [system] [info] Temperatures - Average: 44.64C, CSS: 45.55C, MSS 44.41C, UPA: 44.64C, DSS: 43.96C
[19443010319B191300] [1.1] [48.960] [system] [info] Cpu Usage - LeonOS 40.97%, LeonRT: 30.85%
Manip FPS: 29.89
Depth Map FPS: 30.00
featuresRight 266
featuresLeft 269

jakaskerl

LorenzoR
What happens if you try with a lighter model? I still think the issue might be what I said above.
Up the setNumFramesPool() for the cameras (to eg. 10) and check if any difference is observed.

Thanks,
Jaka

LorenzoR

I tried setNumFramesPool(10) and it seems to make the execution more robust, but the node still crashes sometimes.
I am not sure I get the reasoning behind the fix: since I am already setting the queue size to 1 and the node to be non blocking, if the inference takes too long, shouldn't frames be discarded? Are they stored in RAM till the node "decides" to process or discard them?
Increasing the pool size won't introduce delays if frames are not processed fast enough (the reason why I am setting the queue size to 1), correct?
Thank you!

jakaskerl

LorenzoR
Try to increase the num frames pool for other nodes as well. Since the pipeline is quite complex, one node waiting for the other (stereo waiting for both left and right) can cause a stall if pools run out of frames. You should get a memory issue if too high of a size is set.

Thanks,
Jaka

LorenzoR

jakaskerl

Thank you for your reply.
I can try that, but my concern is that it's only the YOLO node crashing, everything else keeps working perfectly.
And I still don't get why (as explained in the previous message) increasing the pool would affect the execution when the node queue size is set to 1.

jakaskerl

LorenzoR
Because the cameraa stops producing frames if it run out of pool - frames are not being processed fast enough (end to end) and hold the resources which essentially causes a deadlock.

Thanks,
Jaka