• Hardware Platform (Jetson / GPU) 1080 Ti • DeepStream Version 5.1 (Official Docker Container) • Problem Description: Hi, I’m now woring on a face detection task, and I’m having trouble with the output given by the nvinfer. I exported a pre-trained detection model using pytroch to ONNX model, t…

I think I have solved this problem. The cause of the bug should be: Object bbox takes [left, top, width, height]. However, my input is [x1, y1, x2, y2] and that’s why the output looks so weird. Now everything goes well and I can successfully detect faces in Deepstream. But I’m still wondering, in…

Sorry for the late response, we will do the investigation and have the update soon.

Thank you so much for your reply. I have run the auto engine file, generated from the ONNX model, with the python TensorRT and the result is correct. So, maybe there’s something wrong with the nvinfer preprocessing step. I’m working on it and try to find out why.

can you try with add net-scale-factor=0.00392156862745098 into your [property] group

Yes, I’ve tried that before and the Deepstream nvinfer model cannot detect any object. The input image to the PyTorch model is like: tensor([[[[ 114., 114., 114., ..., 81., 81., 81.], [ 114., 114., 115., ..., 81., 81., 81.], [ 114., 115., 115., ..., 81., …

[image] infinitesamsarax: scale = torch.Tensor([img.shape[1], img.shape[0], img.shape[1], img.shape[0]]) img -= (104, 117, 123) img = img.transpose(2, 0, 1) img = torch.from_numpy(img).unsqueeze(0) img = img.to(device) scale = scale.to(device) I…

Thank you for your reply! Yes, it’s preprocessing the image input. I’ve tried another image with shape [3, 640, 368], so resize here is 1.0 and no OpenCV resize operation is applied. But the result in Deepstream is still not correct. The model input shape should be [N, 3, H, W] in both PyTorch an…

Is the output matrix from deepstream the same as the output from pytorch ?

The shape of the output is ok, but the values of the tensor is not.

have you try option maintain-aspect-ratio=1

Sorry for the late reply. I’ve tried it before. But, unfortunately, it did not work.

Custom detection ONNX model gives wrong outputs using nvinfer with DeepStream 5.1

Accelerated Computing Intelligent Video Analytics DeepStream SDK

mchi September 22, 2021, 4:58am 13

Hi @infinitesamsarax ,
Could you try " 2. [DS5.0GA_Jetson_GPU_Plugin] Dump the Inference Input" in DeepStream SDK FAQ - #9 by mchi to dump the inference image sent to TensorRT enqueue() and check if it match with your standalone sample?

Thanks!

Topic		Replies	Views
Object detection pre-trained model inference issue in deepstream DeepStream SDK tensorrt , jetson-inference , gstreamer , python	50	1442	August 9, 2024
Classifier result on onnx doesn't match Deepstream result DeepStream SDK tensorrt , tensorflow , nvbugs , onnx	34	3983	July 28, 2021
Using custom model in deepstream DeepStream SDK jetson-inference , python , deepstream	39	1107	September 10, 2024
TensorFlow EfficientDet-D0 -> ONNX -> TensorRT converted model fails to run in Deepstream DeepStream SDK deepstream61	7	1227	August 11, 2022
Custom model nvinfer facedetection with wrong output problem DeepStream SDK cuda , jetson-inference , gstreamer , deepstream	2	536	September 6, 2023
Run deepstream with Retiface model but got wrong output with different input shape image DeepStream SDK	18	1977	April 29, 2021
Loss of precision to onnx converter for engine by deepstream 6.3 DeepStream SDK tensorrt , gstreamer , inference-server-triton	30	1029	August 2, 2024
Onnx model on deepstream5.0: Nvinfer error: Could not find NMS layer buffer while parsing DeepStream SDK ssd , pytorch , onnx	14	2816	April 15, 2021
Can't configure DeepStream classifier to give the same softmax outputs as the TRT engine it builds DeepStream SDK deepstream , config	23	1461	January 4, 2024
NvDsInferLayerInfo not giving expected no. of outputs DeepStream SDK	59	2862	July 31, 2020

Custom detection ONNX model gives wrong outputs using nvinfer with DeepStream 5.1

Related topics