Parallel-YOLO

Objective

Vision based perception systems are critical for autonomous-driving vehicle products. The objective of this project is to train a tiny-YOLO based CNN pipeline that can detect objects on composite images derived from multiple camera feeds on a vehicle.

Instructions

If compiling under linux, make sure to either set the CUDNN_PATH environment variable to the path CUDNN is installed to, or extract CUDNN to the CUDA toolkit path.

To enable gflags support, uncomment the line in CMakeLists.txt. In the Visual Studio project, define the macro USE_GFLAGS.

Make sure you have all the python dependencies installed listed in requirements.txt.

The major dependencies are cuDNN, CUDA, OpenCV and Numpy.

To compile with and run with CMake, run the following commands:

bash main.sh

This will create targets in a C++ readable format, create evaluation targets and compile the code.

To train the model, modify the configuration parameters in src/model.cu

input_height = input_width = 416;
in_channels = 3;
batch_size = 1;
num_classes = 15;
num_anchors = 5;
learning_rate = -0.001;
num_images = 12;
epochs = 10000;
ITERS = epochs * num_images;
SAVE_FREQUENCY = 50;

Now run the following

cd build
./train

To evaluate the model, place the evaluation images in $ROOT/eval/images/, create an empty directory as $ROOT/eval/predictions and modify the configuration parameters in src/test.cu

input_height = input_width = 416;
in_channels = 3;
batch_size = 1;
num_classes = 1;
num_anchors = 5;
num_images = 1;

Now run the following from $ROOT/build

./eval

To see predictions on the terminal, run the following from $ROOT/scripts/

python infer_targets.py

The resources at hand currently are insufficient for training this architecture. So, the evaluation file may not give expected results.

Presentation Link : https://docs.google.com/presentation/d/19aTHaOBQJV-aQ_BSsfr9PoLtUurHqWZq5CV35UGhI8E/edit?usp=sharing

Reference Paper

Re-Thinking CNN Frameworks for Time-Sensitive Autonomous-Driving Applications: Addressing an Industrial Challenge