I'm using a SSD [https://github.com/Coldmooon/SSD-on-Custom-Dataset].
After the train and the conversion of the caffe model using the model optimizer I'm running a test using the sample "object_detection_demo_ssd_async" using as input images with various dimensions and different form factors.
As result I obtain more or less the same number of object detected as expected, BUT the positions of the bounding boxes are wrong.
Are you using one of the pre-trained models at the bottom of the GitHub page or are you using the modified VGGNet and training it? Also, can you give me more information by attaching the model files(.caffemodel, .prototxt) and giving the command to convert the modelwith Model Optimizer and the command you used to run the Asynchronous object detection sample with the input image attached?
I'm using the modified VGGNet trained by me. Here I uploaded the files generated by the train: https://www.dropbox.com/s/4xz0y2xjsg1rtym/models.7z?dl=0
To convert the model with model optimizer I used this command:
"python mo_caffe.py --input_model VGG_VOC0712_SSD_300x300_iter_10000.caffemodel --input_proto deploy.prototxt"
and I obtained this:
[ SUCCESS ] Generated IR model.
[ SUCCESS ] XML file: C:\Intel\computer_vision_sdk_2018.2.304\deployment_tools\model_optimizer\.\VGG_VOC0712_SSD_300x300_iter_10000.xml
[ SUCCESS ] BIN file: C:\Intel\computer_vision_sdk_2018.2.304\deployment_tools\model_optimizer\.\VGG_VOC0712_SSD_300x300_iter_10000.bin
[ SUCCESS ] Total execution time: 70.12 seconds.
I converted your files a bit differently by just renaming the .caffemodel to be deploy.caffemodel and converted the model by:
sudo python3 mo.py --input_model ~/Downloads/models/deploy.caffemodel
Then I ran the sample and got the proper results upon running the sample with the car image that's in the demo directory. I am on the latest version of OpenVINO. I highly recommend you upgrade to the latest version of OpenVINO and see if the problem persist.
I am curious about how much your demo inference time is. In my case, I use the model your provide with FP16 ,NCS2 cost 500+ms.
Dose it normal? And I am on the latest version of OpenVINO(2019R1).