c++ tensorrt boxes 全是0

[hihiyo]

cuda12.4 tensorrt10.6 在python环境下面适应tensorrt推理结果没有问题，但是在c++环境推理时labels和scores有值但是boxes全是0，推理代码

(cudaMemcpyAsync(buffers[0], input, 1 * 3 * 640 * 640 * sizeof(float), cudaMemcpyHostToDevice, stream));
(cudaMemcpyAsync(buffers[1], input, 1 * 2 * sizeof(int64), cudaMemcpyHostToDevice, stream));
cudaStreamSynchronize(stream);
context->setTensorAddress(engine->getIOTensorName(0), buffers[0]);
context->setTensorAddress(engine->getIOTensorName(1), buffers[1]);
context->setTensorAddress(engine->getIOTensorName(2), buffers[2]);
context->setTensorAddress(engine->getIOTensorName(3), buffers[3]);
context->setTensorAddress(engine->getIOTensorName(4), buffers[4]);
//context->enqueueV3(stream);
context->executeV2(buffers);
(cudaMemcpyAsync(labels, buffers[2], 1 * 300 * sizeof(int64), cudaMemcpyDeviceToHost, stream));
(cudaMemcpyAsync(boxes, buffers[3], 1 * 300 * 4 * sizeof(float), cudaMemcpyDeviceToHost, stream));
(cudaMemcpyAsync(scores, buffers[4], 1 * 300 * sizeof(float), cudaMemcpyDeviceToHost, stream));
cudaStreamSynchronize(stream);

推理后处理

for (int i = 0; i < 300; i++) {
float score =(scores[b * 300 + i ]);
int64 label =(labels[b * 300 + i ]);
if (score < 0.4) {
continue;
}
std::cout << " score " << score << std::endl;
std::cout << " label " << label << std::endl;
//float(&box)[4] = * reinterpret_cast<float(*)[4]>(&boxes[b * 300 * 4 + i *4]);
float x = boxes[b * 300 * 4 + i + 0];
float y = boxes[b * 300 * 4 + i + 1];
float x2 = boxes[b * 300 * 4 + i + 2];
float y2 = boxes[b * 300 * 4 + i + 3];
std::cout << " x " << boxes[b * 300 * 4 + i + 0] << std::endl;
std::cout << " y " << boxes[b * 300 * 4 + i + 1] << std::endl;
std::cout << " x2 " << boxes[b * 300 * 4 + i + 2] << std::endl;
std::cout << " y2 " << boxes[b * 300 * 4 + i + 3] << std::endl;
cv::rectangle(img_resize, cv::Rect(cv::Point(int(x), int(y)), cv::Point(int(x2), int(y2))), cv::Scalar(00, 0, 255), 1);
}

结果打印

score 0.77002
label 0
x 0
y 0
x2 0
y2 0
score 0.491455
label 1
x 0
y 0
x2 0
y2 0

求解

[yuhyun00]

Hi, I think I can see what's causing this.

Looking at your code, both buffers[0] and buffers[1] are being filled from the same input pointer:

cudaMemcpyAsync(buffers[0], input, 13640640sizeof(float), ...); // images
cudaMemcpyAsync(buffers[1], input, 12sizeof(int64), ...); // orig_target_sizes

buffers[1] corresponds to orig_target_sizes, which should hold the original image's height and width as int64 values (shape [1,2]) — it's used inside the postprocessor to rescale the normalized box coordinates back to the original image size. Right now it's being filled with the raw image pixel data (reinterpreted as int64) instead of the actual original size.

Since labels/scores don't depend on orig_target_sizes, they still come out correctly, but the box coordinates get scaled using garbage values, which would explain why they end up as 0.

Try creating a separate host array like:

int64_t orig_size[2] = { orig_h, orig_w };
cudaMemcpyAsync(buffers[1], orig_size, 12sizeof(int64_t), cudaMemcpyHostToDevice, stream);

and see if that fixes the box output.