search

openvinotoolkit / open_model_zoo

Pre-trained Deep Learning models and demos (high quality and extremely fast)
https://docs.openvino.ai/latest/model_zoo.html
Apache License 2.0
4.07k stars 1.37k forks source link

[object_detection_demo_yolov3_async.py] chopped bounding box predictions #337

Closed anhminh3105 closed 5 years ago

anhminh3105 commented 5 years ago

Hello,

I followed the tutorial on https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html to create the IR model graph and used it to run the python demo https://github.com/opencv/open_model_zoo/blob/master/demos/python_demos/object_detection_demo_yolov3_async/object_detection_demo_yolov3_async.py. However, I am experiencing the following strange outputs. The detected bounding boxes are chopped to small overlapping ones, and sometimes the boxes are only a tiny square on that object. I am using OpenVINO version 2019.2.242, TF1.12 and TF1.14, and Ubuntu 18.04. Below is the windows snip from the demo, please help Thanks in advance Screenshot from 2019-08-26 11-11-16

IRDonch commented 5 years ago

Which of the YOLO models are you using? I tried the demo with the full YOLOv3 model, and the bounding boxes were reasonable.

anhminh3105 commented 5 years ago

I was able to solve the problem by regenerating the frozen model and the IR model graph using the OpenVINO version 2019.2.275. Thanks a lot