TDA2EVM5777: [TIDL] Queries about caffe-jacinto layers

Taro Shibata

Part Number: TDA2EVM5777

Hi,

This is regarding TIDL (REL.TIDL.01.02.00.00)

I’m looking caffe-jacinto model.

I would like to know the details about “DetectionOutput” layer and “PriorBox” layer.

- What is done in these layer?
- The meaning of each parameters (min_size, aspect_ratio, share_location, code_type, etc.)

I checked TIDeepLearningLibrary_UserGuide.pdf as well as API guide TIDeepLearningLibrary_API.chm, but could not find good explanation.

Are there any information?

Regards,

-Shibata

0 Manu Mathew over 5 years ago

Hi,

TIDL implements Single Shot Multi Box Detector (SSD). https://arxiv.org/abs/1512.02325

The layers that you mentioned were originally developed by Wei Liu, who implemented SSD in Caffe. You can see the original implementation of SSD here: https://github.com/weiliu89/caffe/tree/ssd

The original implementation of the specific layers that you mentioned are:

You can go through these files to get a better idea.

[caffe-jacinto](https://github.com/tidsp/caffe-jacinto) is derived from [nvidia-caffe](https://github.com/NVIDIA/caffe), the SSD part of which is which in turn inspired by Wei Lu's implementation.

Best regards,

0 Taro Shibata over 5 years ago in reply to Manu Mathew

Hi Mathew,

Thanks for sharing the information.

Please close this thread.

Regards,

-Shibata

Processors forum