Evaluating Transfer Performance of Self Supervised Models for Vehicle Re-Identification

Abstract

In recent times, self-supervised visual representation learning has made considerable advancements. But no such studies are available which determine the extent to which features learned through self-supervised pre-training on the ImageNet dataset can generalize to distinct tasks. Labelled data is one of the major challenges faced in the field of vehicle re-Identification. So, we have decided to perform this analytical study on the downstream task of Vehicle Re-Identification. In our model, architecture is based on two steps to address this problem and assess the transfer performance of SSL method on Vehicle Re-identification task. The first step is to pretrain the self-supervised model, which in our case SimCLRv-2, on ImageNet to learn representations from this data. The pre-training phase leverages unlabeled data to initialize the model and learn representations. The second step is to use these pre trained weights from the intial step to perform vehicle re-identification task. In this step, we have used fully connected convolutional layers on top of the frozen features that are acquired from pretrained model. We have used Veri-WILD dataset to further tarin and test the model. We have experimented with pretraining and without pretraining the model. Our experiments provide evidence that the inclusion of self-supervised pre-training results in quicker convergence and less training time in comparison to training without pre-training. The evaluation metric that we have used is mean Average Precision (mAP%). This strategy has helped us achieve more accurate results, by using both learned representations as well as a few annotations present in the data.