TDA2: Model import and inference on TDA2XX

Hi,

To add, the model was trained using Caffe Jacinto framework.

Please find additional attachments as follows:

1. Cat_Dog_Inference_Desktop.mp4  [Model inference using Caffe on the desktop machine.]
2. Cat_Dog_SSD_Inference.jpg [Model inference using TIDL Inference on TDA2XX ]

Regards,

Premchand

Please find attached folder for previous post.

Cat_Dog_Inference.zip

Hi,

Could you try below..
1. Can you display only strong objects from output, that is objects with score >0.3 only and check the output display
2. Can you change "quantRoundAdd" to 50 (default value) in the import config and check

Thanks,
Praveen

Hi Praveen,

1. I have already been displaying strong objects i.e. score > 0.3.

2. I changed "quantRoundAdd" to 50 in import config. This did not resolve issue on TDA2X however, I observed that detected bounding boxes with visualization code are within the limit of the image of resolution.

Additionally, I tried out few more combinations in import config today which are as follows:

• "quantRoundAdd" to 50 and "numParamBits" to 10.
• "quantRoundAdd" to 50 and "numParamBits" to 8.
• Configured layers with size less than 64 X 64 to have dense convolution.

Neither of these combinations resolved my issue.

Regards,

Premchand 

Hi,

Is there any update on this?

Regards,
Premchand

Can you try setting below create time parameter
createParams.quantMargin = 30;

Hi,

It did not resolve the issue. However, I observed that earlier random false detections were for Dog and now it is for Cat. 

FYI, I had set parameters as follows:

quantHistoryParam1 = 20

quantHistoryParam2 = 10

quantMargin = 30

Please find attached image for detection results.

Is the network produces expeted result during import step for possitive sample?
Is the network produces expeted result in SDK usecase for possitive sample?
Is this issue specific to false dtection in negative smaple?

Hi Kumar,

1. The network produces somewhat positive result to positive sample during the import step. Please find below output for the same.

2. Network does not produce expected result in SDKuse case for a positive sample. Please find below output for the same.

3. The issue is not specific to just negative samples on TDA2XX, issue is also regarding reasonable detections for positive samples as well.

Can you try this cat image with import step and confirm the behavior.

Hi Kumar,

I tried this cat image with import step but the detection is not similar to detection for cat image on TDA2XX.

Please find below image.

Legend for rectangle color:

1. Green -- Dog
2. Red -- Cat

Regards,

Premchand