ConvNet-based visual place recognition under appearance changes for unmanned vehicles

Abstract

Identifying the place unmanned vehicles have visited is crucial for their re-localization to eliminate accumulating drifts. As a VPR(visual place recognition problem), the goal is to retrieve the correct reference frames in database which depict the same place as given query image.For retrieval-basedVPR, methods can be classified as VLAD-based and sum-based. In this dissertation, I present the following contributions. FirstlyI reproduced pipeline of the algorithm, and then trained the modelswhose backbone are alexnet or VGG16 and head architecture are max, avg or VLAD-based pooling layer NetVLAD respectively on the Pittsburgh 30k training setand test them on Pittsburgh 120k test and Pittsburgh 30k val. Then, I evaluatedtheirabilitiesof generalizationby applying them on the revisited image retrieval testing datasets roxford5k and rparis6k. And byintroducing indicator mAP and mP@, their overall performancesarebetter evaluated and compared. What’s more, I reproduced another sum-basedpooling layer APANet, then trained and evaluatedits performance. Finally I showedthat NetVLAD possesses the overall best performance, APANet enjoys greater improvement compared to the sumpooling.