Detectned_V2 bbox

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Description

I trained Detectned_V2 in Transfer Learning Toolkit on kitti dataset (It was very comfortable).

Now i want to use this model in TensorRT(Not in deepstream) in my own c++ aplication.
TensorRT engine has 2 outputs: output_cov/Sigmoid and output_bbox/BiasAdd.
I want to get probability and bounding boxes.
As I understand for probability i must use some variant of DBSCAN algorithm for output_cov/Sigmoid and NMS for output_bbox/BiasAdd
Is there some C++ code sources for postprocessing this outputs (especially for output_cov/Sigmoid postprocessing)?
Another information that i cant find - Is it necessary for this model to use some preprocessing (using mean value)?

Environment

nvcr.io/nvidia/tlt-streamanalytics:v2.0_dp_py2

2

Please refer to below sample if in case it helps:

Thanks

4

I successfully received output_cov/Sigmoid
But i cant receive right bboxes
top- my boxes(all predicted box fromo utput_bbox/BiasAdd ); bot - TLT result

image
image1248Ă—764 738 KB

CODE:
outputIndex1 = engine.getBindingIndex(“output_bbox/BiasAdd”);
float box3[4 * 24 * 78];
cudaMemcpyAsync(box1, buffers[outputIndex1], 4 * 24 * 78 * sizeof(float), cudaMemcpyDeviceToHost, stream);
cudaStreamSynchronize(stream);

///Postprocessing

for (int i = 0; i < 24 * 78; ++i)
{
int dx = (i % 78) ;
int dy = (i / 78) ;
float x1 = (box1[i * 4 + 0] + dx) * 16;
float y1 = (box1[i * 4 + 1] + dy) * 16;
float x2 = (box1[i * 4 + 2] + dx) * 16;
float y2 = (box1[i * 4 + 3] + dy) * 16;
cv::rectangle(image, cv::Point(x1, y1), cv::Point(x2, y2), cv::Scalar(0, 255, 0), 1, 8, 0);
}
Something wrong in my output_bbox/BiasAdd postprocessing.

Regards.

5

Please refer to below sample:

Thanks

6

Hi mazitov

Have you found the proper way to postprocess the output? I would be very apperciated if you can share your code.

Regards

7

#include <src/cnn/nvdsinfer.h>
#include <src/cnn/nvdsinfer_dbscan.h>

void Cnn::assignClass(std::vector &objectList, std::vector &objectListRes,
int num)
{
int objNum = objectList.size();
if (objNum < 1)
return;

for (int i = 0; i < num; ++i)
{
objectListRes[i].classId = objectList[0].classId;
float ax = objectListRes[i].left + objectListRes[i].width / 2;
float ay = objectListRes[i].top + objectListRes[i].height / 2;

float bx = objectList[0].left + objectList[0].width / 2;
float by = objectList[0].top + objectList[0].height / 2;

float dist = sqrt((ax - bx) * (ax - bx) + (ay - by) * (ay - by));

for (int j = 1; j < objNum; ++j)
{
bx = objectList[j].left + objectList[j].width / 2;
by = objectList[j].top + objectList[j].height / 2;

float distItr = sqrt((ax - bx) * (ax - bx) + (ay - by) * (ay - by));
if (dist > distItr)
{
dist = distItr;
objectListRes[i].classId = objectList[j].classId;
}

}
}
}

void Cnn::doInference(IExecutionContext& context, float* inputData, cudaStream_t &stream)
{

cudaMemcpyAsync(buffers[inputIndex0], inputData, batch_size * INPUT_D * sizeof(float), cudaMemcpyHostToDevice, stream);
context.enqueue(batch_size, buffers, stream, nullptr);

float probs[classNum * OUT_DIM_H * OUT_DIM_W];
float boxs[classNum * 4 * OUT_DIM_H * OUT_DIM_W];

cudaMemcpyAsync(probs, buffers[outputIndex0], classNum * OUT_DIM_H * OUT_DIM_W * sizeof(float), cudaMemcpyDeviceToHost, stream);
cudaMemcpyAsync(boxs, buffers[outputIndex1], classNum * 4 * OUT_DIM_H * OUT_DIM_W * sizeof(float), cudaMemcpyDeviceToHost, stream);
cudaStreamSynchronize(stream);

std::vector objectList;

int gridW = OUT_DIM_W;
int gridH = OUT_DIM_H;
int gridSize = gridW * gridH;
float gcCentersX[gridW];
float gcCentersY[gridH];
float bboxNormX = 35.0;
float bboxNormY = 35.0;
float *outputCovBuf = (float *) probs;
float *outputBboxBuf = (float *) boxs;

int strideX = DIVIDE_AND_ROUND_UP(INPUT_W, gridW);
int strideY = DIVIDE_AND_ROUND_UP(INPUT_H, gridH);

for (int i = 0; i < gridW; i++)
{
gcCentersX[i] = (float) (i * strideX + 0.5);
gcCentersX[i] /= (float) bboxNormX;

}
for (int i = 0; i < gridH; i++)
{
gcCentersY[i] = (float) (i * strideY + 0.5);
gcCentersY[i] /= (float) bboxNormY;
}

for (int c = 0; c < classNum; c++)
{
float *outputX1 = outputBboxBuf + (c * 4 * gridW * gridH);
float *outputY1 = outputX1 + gridSize;
float *outputX2 = outputY1 + gridSize;
float *outputY2 = outputX2 + gridSize;

float threshold = 0.02;

for (int h = 0; h < gridH; h++)
{
for (int w = 0; w < gridW; w++)
{
int i = w + h * gridW;
if (outputCovBuf[c * gridSize + i] >= threshold)
{

NvDsInferObjectDetectionInfo object;
object.classId = c;
object.detectionConfidence = outputCovBuf[c * gridSize + i];

float rectX1f, rectY1f, rectX2f, rectY2f;

rectX1f = (outputX1[w + h * gridW] - gcCentersX[w]) * -bboxNormX;
rectY1f = (outputY1[w + h * gridW] - gcCentersY[h]) * -bboxNormY;
rectX2f = (outputX2[w + h * gridW] + gcCentersX[w]) * bboxNormX;
rectY2f = (outputY2[w + h * gridW] + gcCentersY[h]) * bboxNormY;

/* Clip object box co-ordinates to network resolution */
object.left = CLIP(rectX1f, 0, INPUT_W - 1);
object.top = CLIP(rectY1f, 0, INPUT_H - 1);
object.width = CLIP(rectX2f, 0, INPUT_W - 1) - object.left + 1;
object.height = CLIP(rectY2f, 0, INPUT_H - 1) - object.top + 1;

objectList.push_back(object);

}
}
}
}

size_t numObjects = objectList.size();
auto unclasteredObjectList = objectList;
NvDsInferDBScanCluster(DBScan, &DBScanParams, &objectList[0], &numObjects);
assignClass(unclasteredObjectList, objectList, numObjects);

for (int i = 0; i < numObjects; ++i)
{
auto object = objectList[i];
cv::Scalar color(255, 255, 0);
if (object.classId == 0)
color = cv::Scalar(0, 255, 0);
else if (object.classId == 1)
color = cv::Scalar(255, 0, 0);
else if (object.classId == 2)
color = cv::Scalar(0, 0, 255);

cv::rectangle(visualize, cv::Rect(object.left, object.top, object.width, object.height), color, 2, 8, 0);
}
}

1 Like
9

Hi mazitov

Thanks for sharing your code.
Where is the source code of this method

NvDsInferDBScanCluster

I did found the header file in DeepStream SDK, but I can’t found the source code