Openvino performance a lot worse than darknet model

[Jucjiaswiss]

[Required] Your device (RaspberryPi3, LaptopPC, or other device name):
LaptopPC [Required] Your device's CPU architecture (armv7l, x86_64, or other architecture name):
x86_64 [Required] Your OS (Raspbian, Ubuntu1604, or other os name):
Ubuntu1804 [Required] Details of the work you did before the problem occurred:

python3 convert_weights_pb.py \ --class_names yolov3-tiny-mine.names \ --weights_file weights/yolov3-tiny-mine_final.weights \ --data_format NHWC \ --tiny True \ --output_graph pbmodels/frozen_yolov3-tiny-tw.pb
python3 /opt/intel/openvino_2019.1.094/deployment_tools/model_optimizer/mo_tf.py \ --input_model models/tf/frozen_yolov3-tiny-tw.pb \ --output_dir models/openvino/FP32 \ --data_type FP32 \ --batch 1 \ --tensorflow_use_custom_operations_config yolo_v3_tiny_changed.json
[Required] Error message:

performance is a lot worse than darknet inference
[Required] Overview of problems and questions:

I trained the tiny-yolov3 model with darknet and test video perforamnce, it was pretty good. When I conveted it to openvino, the result was pretty bad, it missed several objects. Is it normal? How can I improve?

[pra-dan]

The v3-tiny will show poor performance in detecting all the objects correctly, but its meant for speed. You will surely get faster results; a higher FPS.

[andeyeluguo]

all people meet this bug and nobody can solve it even the intel coporation by now.

[andeyeluguo]

it may decreased 20 MAP