Object tracking project

[jnebrera]

Hi all,

I just purchased an Adlink Vizi-AI board to start experimenting with computer vision as my first pet project. I have opted for this board instead of a Jetson Nano as the documentation for OpenVINO is quite good, and you provide real examples for the Gstreamer side, in particular the DL Streamer component.

Now, my project is about tracking a basketball in a dual Full HD video stream (3.840x1.080) and cropping it to a final Full HD resolution (1.920x1.080) based on ball position.

Now I wonder if the available object tracking capability will be enough for my needs, as it seems it runs in CPU only, and this board has quite a limited one (Intel Atom® x5-E3940 SOC). Actually, I purchased this board because of its Movidius Myriad X on board, and seems surprising now I can't exploit it for this complex task.

Some questions then:

1) Can you please confirm gvatrack runs in CPU mode only? What level of fps can be expected from it?

2) As the task is to track a single object, anybody has ported algorithms like siamrpn, Tracknet or SiamMask to OpenVINO and indirectly to DL Streamer? I have seen some person checking for this in the past, but he didn't get it to work.

Thank you

[Rizal_Intel]

Hi Jnebrera,

Yes, the gvatrack runs in CPU mode only as it has no arguments for devices. You could use the movidius for object detection inference with frame skipping (as inferencing should be the compute-intensive part of the pipeline) and gvatrack on CPU. The FPS could vary according to your pipeline and inference model, but the performance should be similar to Neural Compute Stick 2 as both has single Myriad X VPU.

The conversion of the algorithms you mentioned is not validated yet, but as all of them seems to be implemented using neural network, you could try converting them using general guidelines. Please note that this is not guaranteed to work.

SiamRPN and SiamMask (Pytorch implementation) could be converted from Pytorch model to ONNX format and then converted to OpenVINO IR.

You could load Tracknet (Keras implementation) into Tensorflow 2 framework and export the model in SavedModel format, which the you could try to convert it to IR.

Regards,

Rizal

[Rizal_Intel]

Hi Jnebrera,

Intel will no longer monitor this thread since we have provided a solution. If you need any additional information from Intel, please submit a new question.

Regards,

Rizal