TDA4VE-Q1: monodel 3D detection model decreased a lot on the TDA4VE platform

Hi,

What was the original device you tested on?  There are various capabilities in the TDA4 devices, so if you went from a TDA4VH to a TDA4V,E there will be some performance decline.   

Regards,

Chris

We test all theses  on PC simulation .   We are going to run it on TDA4V,E,but not yet .  We changed the backbone , but it did not work . Our 3d  AP decreaced from 57 of onnx to 20 of  int16 model .  Our aos AP decreased from 40 of onnx to 17 of int16 model . So ,it decreased a lot ,  and unacceptable .

The link is not available. Please retry uploading.

Regards,

Adam

/home/invo/2TB/tidl/edgeai-tidl-tools-09_01_08_00_beiqi/workspace/3dOD/data/data/onnx/300wcamera_alldata_mobile_remap100w_ti_relu6_last.onnx

Sorry， I  upload it again ,but it still turns to error .  I  sent it to you on WeChat .

/cfs-file/__key/communityserver-discussions-components-files/791/300wcamera_5F00_alldata_5F00_mobile_5F00_remap100w_5F00_ti_5F00_relu6_5F00_last.onnx

Thanks for your file. I upload this for you.

Regards,

Adam

Thanks 

Hi 

I am able to reproduce your issue on pc with edgeai tidl tools 10.01.00.02

my onnx result for output0:

/cfs-file/__key/communityserver-discussions-components-files/791/mono3d0.npy

my tidl32 for output0:

/cfs-file/__key/communityserver-discussions-components-files/791/mono3d0.npy

error:

>>> res_0=onnx_0-tidl32_0
>>> res_0
array([[[[ 0.19376373,  1.362814  ,  2.067832  , ...,  1.93466   ,
           1.9272346 ,  1.3495979 ],
         [ 0.20708942,  2.2395124 ,  3.011363  , ...,  1.603961  ,
           1.434555  ,  1.3182487 ],
         [-0.08019924,  0.98118496,  2.0584078 , ...,  0.50134325,
           0.7411618 ,  1.2766123 ],
         ...,
         [-0.20792675, -0.40433073, -0.25784016, ..., -0.24171877,
          -0.15578222,  0.1526947 ],
         [-0.11457157, -0.29353476, -0.40832233, ..., -0.19024849,
           0.01263475,  0.3563528 ],
         [-0.06742382,  0.19369984,  0.2898035 , ...,  0.31108952,
           0.31078148,  0.3159256 ]],

It takes weeks for TI internal to debug tidl bugs. So to help workaround this issue, here are some suggestions:

1. As I recall, your team has a workable mono3d model last year when I work with Miss Yin. Please check with her to find the difference between your model and hers. 

2. You can try replace clip with batchnorm in your model. Conv+ batchnorm always produce better accuracy results.

Chris Tsongas

Please file a jira for this and I think it is a tool issue.

Regard,

Adam