Hi was using yolov5m custom model trained on crowdhuman dataset.
i tried to convert the model IR format using utralytics export.py code and used detect.py there i got the good results.
The same i followed the dlstreamer guide (Yolov5 Model Preparation Example — Intel® Deep Learning Streamer (Intel® DL Streamer) documentation) to convert the model to IR format and used the below model_proc file there i got the multiple and improper bboxes compared to detect.py.
please find the below pipeline and model repo link
Using openvino 2022.3
Model repo: deepakcrk/yolov5-crowdhuman: Head and Person detection using yolov5. Detection from crowd. (github.com)
MODEL_PATH=../models/yolov5_head_openvino_model/FP32/yolov5m.xml
PROC_PATH=../models/yolov5_head_openvino_model/crowdhuman_yolov5.json
cmd="filesrc location=../videos/NVR_ch18.mp4 ! decodebin \
! videoconvert ! video/x-raw, format=BGR \
! queue ! gvadetect model=$MODEL_PATH model-proc=$PROC_PATH threshold=0.7 device=CPU \
! queue ! gvatrack \
! queue ! gvawatermark ! queue ! videoconvert ! fpsdisplaysink sync=false"
gst-launch-1.0 $cmd
{
"json_schema_version": "2.2.0",
"input_preproc": [
{
"format": "image",
"layer_name": "images",
"params": {
"resize": "aspect-ratio",
"color_space": "BGR",
"reverse_channels": true
}
}
],
"output_postproc": [
{
"converter": "yolo_v5",
"output_sigmoid_activation": true,
"do_cls_softmax": true,
"iou_threshold": 0.20,
"bbox_number_on_cell": 3,
"cells_number": 20,
"classes": 2,
"labels": ["person", "head"],
"anchors": [
10.0,
13.0,
16.0,
30.0,
33.0,
23.0,
30.0,
61.0,
62.0,
45.0,
59.0,
119.0,
116.0,
90.0,
156.0,
198.0,
373.0,
326.0
],
"masks": [
6,
7,
8,
3,
4,
5,
0,
1,
2
]
}
]
}
Using dlstreamer pipeline
`dlstreamer.png`
Link Copied
Hi Shekarneo,
Thanks for your patience.
I've run your PyTorch model with detect.py from the GitHub repository, and the inference result was good:
https://github.com/deepakcrk/yolov5-crowdhuman/tree/master
Inference Result of PyTorch model with detect.py:
Before converting your PyTorch model into Intermediate Representation, I've converted your PyTorch model into an ONNX model. Are you able to run the ONNX model with your detect.py? I would like to check the inference result of the ONNX model with your detect.py before we convert it into an Intermediate Representation.
Regards,
Wan
Hi Shekarneo,
Thanks for the information.
I ran the detect.py script with the command: python3.9 detect.py --weights crowdhuman_yolov5m.onnx --source data/images/bus.jpg --imgsz 640. However, I encountered the following error: detect.py: error: unrecognized arguments: --imgsz 640
Next, I ran the detect.py script without --imgsz 640, I encountered the following error: _pickle.UnpicklingError: invalid load key, '\x08'
Could you please share the ONNX model with us so we can replicate the issue on our end?
Regards,
Wan
Hi Shekarneo,
Thanks for your patience.
I also encountered an accuracy issue when running the object detection pipeline using YOLOv5m with Intel® Deep Learning Streamer Pipeline Framework Release 2023.0.
1. Object Detection using PyTorch model with detect.py
python3 detect.py --weights ../crowdhuman_yolov5m.pt --source ../people-detection.mp4 --device CPU
2. Object detection using Intermediate Representation with Intel® DL Streamer
gst-launch-1.0 filesrc location=people-detection.mp4 ! decodebin force-sw-decoders=true ! queue ! gvadetect model=crowdhuman_yolov5m.xml model-proc=yolo-v5.json inference_interval=1 threshold=0.4 device=CPU ! queue ! gvawatermark ! videoconvert ! autovideosink sync=false
{
"json_schema_version": "2.2.0",
"input_preproc": [
{
"layer_name": "images",
"format": "image",
"params": {"resize": "aspect-ratio","color_space":"BGR"}
}
],
"output_postproc": [
{
"converter": "yolo_v5",
"output_sigmoid_activation": true,
"do_cls_softmax": true,
"iou_threshold": 0.4,
"classes": 2,
"anchors": [
10.0,
13.0,
16.0,
30.0,
33.0,
23.0,
30.0,
61.0,
62.0,
45.0,
59.0,
119.0,
116.0,
90.0,
156.0,
198.0,
373.0,
326.0
],
"masks": [
6,
7,
8,
3,
4,
5,
0,
1,
2
]
}
]
}
Let me check with relevant team and I'll update you as soon as possible.
Regards,
Wan
