TDA4VM: Using the inYuvFormat parameter to increase the YUV2RGB operation will darken the image display

chao yang

Prodigy 170 points

Part Number: TDA4VM

## Using the inYuvFormat parameter to increase the YUV2RGB operation will darken the image display

------

## Use environment and data description：

**Use environment：SDK8.0 & SDK8.2**

**data：** use_data.tar.gz

**Data description：**

**Model：**
use_data/model/
use_data/model/yuv2rgb_demo.prototxt
use_data/model/yuv2rgb_demo.caffemodel

**Quantize import files**
use_data/tidl_import_demo.txt

**Verify image NV12 (2368x352)**
use_data/input_images/
use_data/input_images/img_2368x352_one_for_test.bin

**Result output**
use_data/out_data/
use_data/out_data/out_data.bin
use_data/out_data/out_data.bmp

**Quantify the SDK8.0 model**
use_data/quant_model/
use_data/quant_model/tidl_io_yuv2rgb_demo_1.bin
use_data/quant_model/tidl_net_yuv2rgb_demo.bin
use_data/quant_model/tidl_net_yuv2rgb_demo.bin.svg

---

## Phenomenon：

Through experimental comparison, it is found that using the inYuvFormat parameter to increase the YUV2RGB operation will make the exported image darker.

--------

## Experimental record：

### Import quantification

1. use_data/tidl_import_demo.txt：

```c
modelType = 0
numParamBits = 8
numFeatureBits = 8
quantizationStyle = 2
inputNetFile = "/media/root/data/tmp/yuv2rgb_demo/yuv2rgb_demo.prototxt"
inputParamsFile = "/media/root/data/tmp/yuv2rgb_demo/yuv2rgb_demo.caffemodel"
outputNetFile = "/media/root/data/tmp/yuv2rgb_demo/out/tidl_net_yuv2rgb_demo.bin"
outputParamsFile = "/media/root/data/tmp/yuv2rgb_demo/out/tidl_io_yuv2rgb_demo_"
inWidth = 2368
inHeight = 352
inNumChannels = 1
inData = "/infer/public/yuv/input_images/img_2368x352_one_for_test.bin"
perfSimConfig = "../../test/testvecs/config/import/device_config.cfg"
postProcType = 1
inFileFormat = 1
rawDataInElementType = 0
foldPreBnConv2D = 1
inYuvFormat = 0

```

Question:

1.How to improve this situation？

2.Where can I find the YUV2RGB formula in the SDK？

over 3 years ago

0 Anshu Jain over 3 years ago

TI__Guru 56820 points

Hi,

I couldn't follow the problem here, do you want to feed yuv 420 input to TIDL convert the same to RGB? Can you also confirm if your network is as expected with TIDL host emulation mode?

Regards,
Anshu

0 chao yang over 3 years ago in reply to Anshu Jain

Prodigy 170 points

Hi, Anshu

Set inYuvFormat = 0

I don't understand where the YUV2RGB operations are implemented.

I understand the process as follows：

I had provided input in the "YUV420 NV12" format, In that case layer0 should have only one channel Y and layer1 should be UV component. Layer2 is the conversion layer and Layer 3 is the concat layer , I found that the third layer（concat[3,3] output） data is not RGB. the third layer I understand the output data format is I420.

I pack each layer of data as layer_output.tar.gz.

layer_output.tar.gz

0 Anshu Jain over 3 years ago in reply to chao yang

TI__Guru 56820 points

Hi,

The actual color conversion operation gets absorbed to the following convolution layer i.e. layer 4 in your case. So layer 4 will do convolution along with operation for color conversion.

Regards,
Anshu

0 chao yang over 3 years ago in reply to Anshu Jain

Prodigy 170 points

Thanks, Anshu.

I have a question about how to verify that there is no loss in joining YUV2RGB operations.

Use the same model, method 1:I used inYuvFormat = 0 and input was "YUV420 NV12" format, method 2 : I don't used inYuvFormat = 0 and input was "RGB" format. I found the results of both methods to be biased.

0 Anshu Jain over 3 years ago in reply to chao yang

TI__Guru 56820 points

Hi,

Sorry I couldn't follow your question, what do you mean by the following :

chao yang said:
I found the results of both methods to be biased.

Regards,

Anshu

0 chao yang over 3 years ago in reply to Anshu Jain

Prodigy 170 points

There is a difference between the two ways in my experiment.

Use the same model, method 1:I used inYuvFormat = 0 and input was "YUV420 NV12" format, method 2 : I don't used inYuvFormat = 0 and input was "RGB" BMP format.

How can I verify that the YUV2RGB image implemented using the parameter inYuvFormat is the same as the BMP image data？ I want to prove that inYuvFormat operation has no loss.

Left: method 1

Right: method 2

I compared the 1x32x480x160 layers，Deviations started from this layer.

The following are the data files for each layer.

layer.tar.gz

0 kumar.desappan over 3 years ago in reply to chao yang

TI__Mastermind 22145 points

I believe these two-layer traces have different tensor sclae. We can not directly compare them.

Can you check the tensor scale of these two layers?

BTW, because of quantized computation layer traces may not exactly match. Did you try to generate an accuracy metric for the final output of the model and compare it?