Introduction

When using OpenCV’s Deep Neural Network (DNN) module with Compute Unified Device Architecture CUDA, you may encounter the following error:

(-215:Assertion failed) preferableBackend != DNN_BACKEND_CUDA || IS_DNN_CUDA_TARGET(preferableTarget)

This error occurs when OpenCV’s DNN module fails to properly configure CUDA as the preferred backend for inference. In this article, we will explore why this happens, how we fixed it, and why we had to rebuild OpenCV.

---

Understanding the Error

This assertion failure happens in OpenCV when:

• CUDA is selected as the backend ( DNN_BACKEND_CUDA ), but
• The target ( DNN_TARGET_CUDA ) is not properly recognized as a CUDA-compatible target.

The primary causes of this issue include:

1. Missing cuDNN: OpenCV’s DNN module requires cuDNN to run inference on GPUs.
2. Incorrect OpenCV Build: OpenCV must be built with CUDA and cuDNN support.
3. Incompatible CUDA or cuDNN Versions: The installed versions of CUDA and cuDNN must match OpenCV’s build settings.
4. Incorrect Target Selection: The model may need to use DNN_TARGET_CUDA_FP16 instead of DNN_TARGET_CUDA .

---

Steps to Fix the Issue

1. Verify CUDA and cuDNN Installation

Check whether CUDA and cuDNN are installed correctly:

nvcc --version
ls /usr/local/cuda/lib64 | grep cudnn

If cuDNN is missing, install it by following the steps in the next section.

2. Install cuDNN

Download and install cuDNN 8.9.7:

tar -xvf cudnn-linux-x86_64-8.9.7_cuda11-archive.tar.xz
cd cudnn-linux-x86_64-8.9.7_cuda11-archive
sudo cp -P include/cudnn*.h /usr/local/cuda/include/
sudo cp -P lib/libcudnn* /usr/local/cuda/lib64/
sudo chmod a+r /usr/local/cuda/include/cudnn*.h /usr/local/cuda/lib64/libcudnn*

3. Rebuild OpenCV with cuDNN Support

Since OpenCV was not detecting cuDNN, we had to rebuild it with explicit CUDA and cuDNN support:

cd ~/opencv/build
rm -rf *
cmake .. -DOPENCV_EXTRA_MODULES_PATH=../../opencv_contrib/modules \
-DWITH_CUDA=ON \
-DWITH_CUDNN=ON \
-DCUDNN_LIBRARY=/usr/local/cuda/lib64/libcudnn.so \
-DCUDNN_INCLUDE_DIR=/usr/local/cuda/include \
-DCUDA_ARCH_BIN=6.1 \
-DOPENCV_DNN_CUDA=ON \
-DCMAKE_BUILD_TYPE=Release
make -j$(nproc)
sudo make install

4. Verify OpenCV Build

After installation, check if OpenCV detects cuDNN:

python3 -c "import cv2; print(cv2.getBuildInformation())" | grep -i cudnn

If properly installed, you should see output similar to:

cuDNN: YES (ver 8.9.7)

5. Update Model Code

To ensure compatibility, update your model loading code:

model.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);
model.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA_FP16);

After these steps, OpenCV should successfully run with CUDA and cuDNN acceleration.

---

Why We Had to Rebuild OpenCV

1. Default OpenCV Binaries Lack CUDA Support

The OpenCV binaries installed via apt install libopencv-dev are built without CUDA support. This means we had to manually build OpenCV with CUDA and cuDNN.

2. Ensuring Compatibility with CUDA and cuDNN

We needed to ensure that OpenCV was compiled against the correct CUDA and cuDNN versions to avoid conflicts.

3. Optimized DNN Performance

Building OpenCV with CUDA significantly improves the speed of deep learning inference compared to CPU execution.

---

Conclusion

The (-215:Assertion failed) error in OpenCV’s DNN module occurs due to improper CUDA backend configuration. By installing cuDNN, rebuilding OpenCV with CUDA support, and correctly setting the DNN target, we successfully resolved this issue. This process ensures that OpenCV leverages GPU acceleration for deep learning tasks, significantly improving performance.

By following these steps, you can troubleshoot and resolve similar issues when using OpenCV’s DNN module with CUDA. 🚀