OpenVINO custom Yolov3-tiny / Yolov4-tiny Inference

[bartolini188]

I've been trying to inference yolov3-tiny model in OpenVINO format, but I've ran into some problems:

- first I tried with AlexeyAB darknet repository   repository to train yolo, models worked well (about 60-70mAP) but I could not covert (both v3-tiny and v4-tiny) I tried to convert following OpenVINO tutorial on Yolo  using TensorFlow 1.12.1 since:

IMPORTANT: The correct version of tensorflow has to be used, otherwise the conversion goes wrong! 1.12.2 works. >=1.13.0 does not work. See: https://software.intel.com/en-us/forums/computer-vision/topic/807383
then run:

python3 tensorflow-yolo-v3/convert_weights_pb.py --class_names signals.names --data_format NHWC --w

Although IR model was empty,

- then I've used Pjreddie Darknet Repository to train model     again followed OpenVINO tutorial on Yolo  with yolov3-tiny.json file as configurations file and running with TensorFlow 1.12.1 - It succeed with no errors, however when tried to run inference (OpencVINO version: 2021.4.752) does not run as expected - besides true detections I'm receiving false ones at the edges and a lot of missed ones - boars are seen as cows, deers, etc. 

Inference in OpenVINO

Inference in DarknetInference in Darknet

As seen above, there's a lot of error in OpenVINO IR format Inference, I'd like to improve this of course, I've been running this in Python - could somebody please share working piece of code to check if my inference code is proper? I am attatching .bin and .xml files as well as test image and video.

[IntelSupport]

Hi Bartolini188,

Thanks for reaching out.

I am able to convert yolo v3 model by running the converter using the below command:

python3 convert_weights_pb.py --class_names coco.names --data_formatNHWC --weights_fileyolov3-tiny.weights --tiny

python3 convert_weights_pb.py --class_names coco.names --data_format NHWC --weights_file yolov3.weights

You can refer to YOLOv3 TensorFlow* Model for more information. Can you provide me the source of your yolo v3 code?

Apart from that, please share your inference code and all necessary file for further testing on our end. You can share it here or privately to my email:  noor.aznie.syaarriehaahx.binti.baharuddin@intel.com

Regards,

Aznie

[IntelSupport]

Hi Bartolini188,

Thank you for your question. If you need any additional information from Intel, please submit a new question as this thread is no longer being monitored. 

Regards,

Aznie