Algorithm used: RCNN
Dataset: Images with corresponding LiDar data and bounding boxes coordinates, Cifar-10 dataset. The system was set up with MATLAB and CUDA accelerators on Window 10 for training the Convolutional Neural Networks in this project.
All the images were down-sampled and grey scaled for both training and testing(attached extractLearningData.m file which generates the modified training dataset). The bounding boxes were checked for validity to ensure it is visible in the image, not occluded too much, close enough, small enough and doesn’t belong to any class labels such as Unknown, Motorcycles, Industrial, Cycles, Boats, Helicopters, Planes, Commercial, Trains. This was done before training the CNN. There are two parts in R-CNN algorithm:
-
Training a basic CNN to detect objects in the images (this is our primary classifier that says whether a given regional proposal contains a car or not). Cifar-10 dataset was downloaded and used for training this CNN.
-
Obtaining the region proposal network (RPN) for generating region proposals and training the RCNN. The R-CNN was trained on around 2100 training images from the dataset given as part of this challenge. (we did this as a lot of images were repetitive and to save computation time).
The regional proposal network ranks region boxes and proposes the region most likely containing the objects. Each of these regions are tested on the trained CNN to determine whether the region contains the cars. We repeat this for all the test images and obtain the number of cars in the image along with ground truth bounding boxes.