DeepStream: YOLOv8 ONNX Model Not Detecting with Multiple Camera Streams

victorroza22 · March 3, 2025, 2:54pm

Hello everyone,

I’m facing a problem when using a YOLOv8 model that I trained on DeepStream.

I have a .pt model that I converted to .onnx, and I noticed the following behavior:

When I use only one camera, the model works normally and detects correctly.

When I increase the number of cameras in the pipeline, the model does not make predictions.

If I set batch-size=1 in nvstreammux, the model works for all cameras, but the performance is very low.

When I increase the batch-size of nvstreammux (for example, for the number of cameras), the processing is fast, but it does not detect anything in the videos.

On the other hand, when I do the same test with a pre-trained person detection model, everything works perfectly, even with multiple cameras and a high batch-size.

Therefore, I suspect that the problem is in the generation of the .onnx from my trained YOLOv8 model.

Has anyone experienced this or could tell me if there is a specific configuration in the .onnx export or in pgie that could cause this problem?

I can share the .pt, .onnx and the configs, if it helps in the analysis.

Thanks!

models .pt and .onnx
https://drive.google.com/drive/folders/12n2EKTjTT244R6KXF4QiKZ8n-k55niDX?usp=sharing

Y-T-G · March 3, 2025, 3:20pm

What’s the export command you used to export to ONNX?

victorroza22 · March 3, 2025, 4:14pm

I used this repository with dynamic batch size:

github.com/marcoslucianops/DeepStream-Yolo

utils/export_yoloV8.py

master

import os
import sys
import onnx
import torch
import torch.nn as nn
from copy import deepcopy

import ultralytics.utils
import ultralytics.models.yolo
import ultralytics.utils.tal as _m

sys.modules['ultralytics.yolo'] = ultralytics.models.yolo
sys.modules['ultralytics.yolo.utils'] = ultralytics.utils


def _dist2bbox(distance, anchor_points, xywh=False, dim=-1):
    lt, rb = distance.chunk(2, dim)
    x1y1 = anchor_points - lt
    x2y2 = anchor_points + rb
    return torch.cat((x1y1, x2y2), dim)

This file has been truncated. show original

yuweiw · March 4, 2025, 1:42am

Could you refer to our deepstream_tools to try to use the yolov8 model we provided?

victorroza22 · March 5, 2025, 11:18pm

apparently this command here worked for multiple cameras still in the repo:

github.com/marcoslucianops/DeepStream-Yolo

utils/export_yoloV8.py

master

import os
import sys
import onnx
import torch
import torch.nn as nn
from copy import deepcopy

import ultralytics.utils
import ultralytics.models.yolo
import ultralytics.utils.tal as _m

sys.modules['ultralytics.yolo'] = ultralytics.models.yolo
sys.modules['ultralytics.yolo.utils'] = ultralytics.utils


def _dist2bbox(distance, anchor_points, xywh=False, dim=-1):
    lt, rb = distance.chunk(2, dim)
    x1y1 = anchor_points - lt
    x2y2 = anchor_points + rb
    return torch.cat((x1y1, x2y2), dim)

This file has been truncated. show original

command:
python3 export_yoloV8.py -w /home/verissimo/DS7.1/ds_analytics/models/yolov8/yolo_best.pt --simplify --dynamicapparently this command here worked for multiple cameras still in the repo:

github.com/marcoslucianops/DeepStream-Yolo

utils/export_yoloV8.py

master

import os
import sys
import onnx
import torch
import torch.nn as nn
from copy import deepcopy

import ultralytics.utils
import ultralytics.models.yolo
import ultralytics.utils.tal as _m

sys.modules['ultralytics.yolo'] = ultralytics.models.yolo
sys.modules['ultralytics.yolo.utils'] = ultralytics.utils


def _dist2bbox(distance, anchor_points, xywh=False, dim=-1):
    lt, rb = distance.chunk(2, dim)
    x1y1 = anchor_points - lt
    x2y2 = anchor_points + rb
    return torch.cat((x1y1, x2y2), dim)

This file has been truncated. show original

command:
python3 export_yoloV8.py -w /home/verissimo/DS7.1/ds_analytics/models/yolov8/yolo_best.pt --simplify --dynamic

victorroza22 · March 5, 2025, 11:20pm

yuweiw · March 6, 2025, 1:33am

Glad to hear that. If there are any other questions, you can file a new topic.

system · March 20, 2025, 1:33am

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.

Topic		Replies	Views
Accuracy issues in object detection models using Jetson AGX Orin DeepStream SDK yolo , deepstream , inception	9	740	December 8, 2023
Possible Solutions to INT64 clamping accuracy drop DeepStream SDK tensorrt	11	538	March 11, 2024
Is there any cpp parser for onnx yolov8? DeepStream SDK deepstream	12	75	October 30, 2024
How Can I use custom yolov9 model on Jetpack 5.1.1(DeepStream 6.2) use Gstreamer detect person? DeepStream SDK deepstream	2	46	November 1, 2024
Multicamera detection issues in deepstream 7.0 with jetson DeepStream SDK deepstream	3	56	February 11, 2025
Discrepancies in Object Detection Results between Ultralytics and DeepStream at Yolov8 DeepStream SDK	5	842	February 26, 2024
YOLOv8-l Segmentation, causes sudden spikes in GPU load, when two streams are parallely inferencing DeepStream SDK tensorrt , cuda , gstreamer	20	671	April 1, 2024
Converting yolov7 to deepstream DeepStream SDK	5	1570	June 9, 2023
Loss of precision to onnx converter for engine by deepstream 6.3 DeepStream SDK tensorrt , gstreamer , inference-server-triton	31	509	August 2, 2024
Deepstream testing: Yolov8 models for multiple streams DeepStream SDK deepstream	5	76	July 1, 2025

DeepStream: YOLOv8 ONNX Model Not Detecting with Multiple Camera Streams

Related topics