2018AICITY_LasPalmas

Unsupervised Vehicle Re-Identification using Triplet Networks

This repository contains the code of track3 in the NVIDIA AI City Challenge at CVPRW 2018, in which the sixth place was obtained. This project was developed by Las Palmas de Gran Canaria University and Modena and Reggio Emilia University.

Introduction

This repository contains our implementation to vehicle re-identification. We use a SSD detector to detect the different vehicles and then, a triplet architecture based on VGG to generate an embedded feature space. After that, a quadruple strategy is used to match the vehicles.

Code structure

1. In data_parser is the code to generate the dataset of one video. After in utils/preProcesedDataSets.py is the code to remove wrong detections.
2. In utils, you have utilities to shuffle the dataset (shuffleDataSet.py), concatenate different datasets generated by the data_parser (concatDataSets.py) and in the case that you have cropped the video, a script to store the true bounding box of the detection (correctBoundingboxDataSet.py).
3. trainTriplet.py generates a model that is used in reduced_dataSet_generator.py to transform the detections to a new feature vector generating a new reduced dataset with only one frame per vehicle ID.
4. The re-id strategy is comprised in sorted_quadrupla_list_generation.py script.
5. To output the results in the format of the challenge, you can use k_top_output-template.py.

Resources

1. Pretrained SSD model. detector_model
2. Trained triplet network model, dimension of 100 to feature embedding. triplet_model

Reference

Pedro A. Marín-Reyes, Andrea Palazzi, Luca Bergamini, Simone Calderara, Javier Lorenzo-Navarro, Rita Cucchiara. Unsupervised Vehicle Re-Identification using Triplet Networks, in Proc. IEEE/CVF International Conference on Computer Vision and Pattern Recognition Workshops (CVPRW). Salt Lake City, EEUU, June 2018.