Model_Info.md

Traffic Light Classifier Models

This document provides an overview of all the models our team used to classify traffic lights.

List of Models

1. faster_rcnn_sim
2. faster_rcnn_reallife
3. ssd_inception_v2_coco_sim
4. ssd_inception_v2_coco_reallife
5. ssd_sim_and_real_24_03_2018
6. ssd_sim_and_real_30_03_2018

Model Details

faster_rcnn_sim

This model is optimized to classify simulator images. This model is based on Faster R-CNN Resnet 101 model architecture and created using TensorFlow's Object Detection API. Model is first trained using Bosch Small traffic Light Data set and then fine tuned for a hand annotated set of images from the Udacity Simulator.

Model Params:

epochs : 10000 learning rate : 0.0003

Loss Graph

Few Inference Images

faster_rcnn_reallife

This model is optimized to classify reallife traffic light images. This model is based on Faster R-CNN Resnet 101 model architecture and created using TensorFlow's Object Detection API. Model is first trained using Bosch Small traffic Light Data set and then fine tuned for a hand annotated set of images from the Udacity test track.

Model Params:

epochs : 10000 learning rate : 0.0003

Loss Graph

Few Inference Images

| |

ssd_inception_v2_coco_sim

This model is designed to classify simulator images. This model is based on SSD inception v2 model architecture and created using TensorFlow's object detection API. It is trained using a hand annotated set of images from the Udacity Simulator.

Model params:

epochs : 5000 learning rate : 0.004

ssd_inception_v2_coco_reallife

This model is designed to classify reallife traffic light images. This model is based on SSD inception v2 model architecture and created using TensorFlow's object detection API. It is trained using a hand annotated set of images from the Udacity test track.

Model params:

epochs : 5000 learning rate : 0.004

ssd_sim_and_real_24_03_2018 ssd_inception_v2_coco

This model is designed to classify both reallife and simulator traffic light images. This model is based on SSD inception v2 model architecture and created using TensorFlow's object detection API. It is trained using a hand annotated set of images from the Udacity test track, as well as using Bosch Small traffic Light Data set

Model params:

epochs : +30k learning rate : 0.004

Loss Graph

Few Inference Images

ssd_sim_and_real_30_03_2018 ssd_inception_v2_coco

This model is designed to classify both reallife and simulator traffic light images. This model is based on SSD inception v2 model architecture and created using TensorFlow's object detection API. It is trained using a hand annotated set of images from the Udacity test track, Bosch Small traffic Light Data set, as well as hand annotated by sloth images

Model params:

epochs : +100k learning rate : 0.004

Loss Graph

Few Inference Images