Quantized YOLOv5x6 for Traffic Object Detection

There are many powerful object detectors proposed up to this day. Those models are mostly targeted to general purpose object detection. However, this is unnecessary in the realm of autonomous driving. In fact, the complexity of the models seriously requires huge computational power from workstation-scale GPUs, rendering the models inapplicable in embedded environments. We attempt to address this issue by performing quantization using TensorRT. Choosing the YOLOv5x6 as the object detection model, the model is trained and tested on the KITTI 2D Object Dataset. The model is then quantized to half-precision floating point. Results show that the model can achieve faster speed when the trained model parameters are quantized, while having the detection mAP similar to the non-quantized version.