How to get axie of detection object form gGraph.GetResult()

[idata]

** when I use the intel movidious to detect object ,the network is ssd_mobilenet of tensorflow,I want to get the axie of detection object .**

I see the stream_infer example of movidious:

1. inferenc_result,user_data=gGraph.GetResult() //but it only get the index of object.

Any one know how to get axie of detection object,use tensorflow modle in Movidius stick?

thank you very much!

[idata]

@Garrett Currently SSD MobileNet is not supported for TensorFlow. I'm not sure what you mean when you say axie of detection, but stream infer is a image classifier so it makes sense for it to return the index of the object. Object detection networks may return the number of detected objects and their location in the image as seen in the ncappzoo's video object app which uses SSD Caffe. https://github.com/movidius/ncappzoo/blob/master/apps/video_objects/video_objects.py#L72.

As you can see from the code, it returns the following items:

0: image_id (always 0 for myriad)

1: class_id (this is an index into labels)2: score (this is the probability for the class)3: box left location within image as number between 0.0 and 1.04: box top location within image as number between 0.0 and 1.05: box right location within image as number between 0.0 and 1.06: box bottom location within image as number between 0.0 and 1.0

[idata]

@Tome_at_Intel Thank you very much for your help. I follow your instruction test the video objects .

in the source code:

I see the labels has " car ,person.motorbike" so on ,but the Movidious can not detect the person and car in the video,why? low recognition rate?

Look forward to your reply

[idata]

@Garrett There are a lot of things that can have an affect on an object detector network like SSD MobileNet. SSD MobileNet actually detects a lot of items but in our appzoo example, they are filtered out with a minimum threshold conditional statement. Lowering and increasing this threshold can affect the number of detected items. https://github.com/movidius/ncappzoo/blob/master/apps/video_objects/video_objects.py#L34. There are other factors that can affect object detection like the dataset you trained the network with, the quality of the videos/images you are using, the preprocessing you are performing/not performing, etc.

[idata]

@Tome_at_Intel Thanks for your reply, it's very useful to me.

By the way, i would like to consult you with another 2 questions:

1) I tried tiny_yolo deep learning model (based on Caffe framework) on Movidius, it does not run smoothly compared with Tensor flow framework, and it has about 1+ second delay. Do you know what's the reason, and how to improve its performance ?2) Besides Tiny_yolo, do you have any others could be referred to me ? How many models supported by Movidius NCS stick, any link or list could be provided ? Any models suit for Top view human detection ?

Thanks again for your kind help, and looking forward to your reply!

[idata]

@Garrett I would recommend SSD MobileNet on Caffe over Tiny Yolo. SSD MobileNet on Caffe can be found on the NCAPPZOO at https://github.com/movidius/ncappzoo/tree/master/caffe/SSD_MobileNet.

We don't have a model that can be used for top view human detection, but it would be possible to re-train SSD MobileNet Caffe using (https://github.com/listenlink/caffe/tree/ssd) if you have a data-set.

[idata]

@ Tome_at_Intel Thanks again.

I want to compile the VGGNet PASCAL VOC models SSD512 caffe modle to produce graph, there are SSD_MobileNet AgeNet GoogLeNet so on in the ncappzoo/caffe/ folder,which similar model can I modify to compile VGGNet modle.？Thanks again

[idata]

@Garrett If you want to do object detection, I recommend using SSD MobileNet on Caffe. If you have a model that you want to try on the NCS, please send me the link and I will be happy to help you test the network.

[idata]

@Tome_at_Intel Is that Movidius only use the modles which in NCS ?

I download the PASCAL VOC models in https://drive.google.com/file/d/0BzKzrI_SkD1_X3ZXQUUtM0xNeEk/viewcompile the models SSD512 to generate graph, but show[Error 35] Setup Error:Not engough resources on Myriad to process this network. Is the Movidius support this modle? so I want to the confirm the avaiable modles about caffe which can be used in Movidius compute stick.Look forward to your favourable reply.

[idata]

@Garrett You can find the models that have been tested with the NCSDK on this page. I think that your network's input is very large (512 x 512) and this may be the reason you are receiving the error [Error 35] Setup Error: Not enough resources on Myriad to process this network.