• State TI Thinks Resolved
  • Locked Locked
  • Replies 1 reply
  • Answers 1 answer
  • Subscribers 98 subscribers
  • Views 799 views
  • Users 0 members are here
Support feedback
Options
Options
Related

This thread has been locked.

If you have a related question, please click the "Ask a related question" button in the top right corner. The newly created question will be automatically linked to this question.

TDA2EVM5777: Object Detection, Vision SDK Next Release Date and TIDL Quantization/Dequantization

SagarK
Intellectual 705 points

Part Number: TDA2EVM5777

Hello,

I have several queries

1. Vision SDK  exact release date

- When TI is planning to release Vision SDK with Reshape Layer support and SSD Object detection example running on it ? Kindly provide date.

2. Variable Input Size for TIDL ?

- I have small object detection network with all conv layers. On caffe , I can give any input size and I will get output accordingly. But, In TIDL it accepts fixed input size. Is it possible to add this feature to TIDL ?

- e.g.  In caffe,  1x3x480x640  ---- > 1x2x235x315  or 1x3x240x320 ---> 1x2x115x155 

In TIDL, Once the input size is fix. It will give the same output. I am not able to provide variable input size.

3. TIDL Quantization / Dequantization

- I am referring to quantization code in tidl_import_common.cpp

- If i want to de-quantize the output from TIDL to get floating outputs,  I have only (quantPrec*256 ) from quantStatTool for each layer. But, How to get offset value or min, max range used while quantization .

- Please Guide us.

4. Which is better for DSP or EVE to run object detection model ?

- EVE or DSP, Can we distribute the single model task among DSP and EVE

5. In object detection model, There will be localization as an extra output along with classification

- localization is nothing but regression. Shall we go for DSP for this task? 

- Quantization will introduce lot of error in the regression output, which will give wrong localization

- How TI is planning to solve this issue? Your thoughts

6. Input Image Preprocessing

- All the preprocessing codes have uint8 output. But, what if we have to normalize the input image  something like this (img - 125)  / 125 .

Thanks in advance

Sagar

 

  • 0
    TI__Mastermind 22145 points

    Hi Sagar,

    1. The Vision SKD release with SSD support is planned for March End. Reshape will be only supported in the context of SSD. It will not be supported as separate layer.

    2. Input size is fixed at create time runtime modification is will not be supported. Multiple algorithm instances with variable size cane created to handle multiple resolutions

    3 Q value will be part of out args. 

    4. We can split network between EVE and DSP.

    5 Please wait for release with SSD support understand this Question.(3 and 4 can be understood well).

    6. Pre-processing 2 and 3 use int8 as input. FYI : meanVlaues is subtracted from the input 

    Regards,

    Kumar.D