objectDetection.md

Object Detection

These sections include a breif description of the usage of the code for the vehicle detection implementation. The CNN based implementation has three heads used for classification, bounding boxes and range estimation.

The implementation uses a pre-trained version of the Resnet-34, available here. Be sure to set the correct path to the pre-trained model inside the function initialiseModel() in model_resnet.lua.

Data set

The network is trained usig the KITTI object detection data set, and uses multiple threads (donkeys) to load the images from disk with functions declared in donkeyCrops.lua. Make sure to set the correct path to the data set:

local dataPath = '/......./KITTI_Object_Detection/'

Same goes for the global opt.path variable, declared in main.lua.

Training

The script main.lua contains useful parameters for the network training, stored in table opt. It also declares a high-level training function train(). The parameter opt.criterionWeights is used to balance the cost functions used for the three heads and can be set to 0 if any of them should be ignored during training. To speciy if gradients should propagate through the entire network or only the heads of the network, use the function setParameterNetwork() with arguments 'model' or 'heads'.

Testing

test.lua contains a set of functions used for testing a trained network qualitatively.