A Strong Baseline for Tiger Re-ID and its Bag of Tricks

As an instance-level recognition task, person re-identification methods always calculate local features by horizontal pooling. It is based on a simple assumption that pedestrians always stand vertically. But as to wildlife re-identification task, we can not make similar assumption since the various view-angles of wildlife. In this paper, we propose a novel dynamic partial matching method. In our module, global feature learning benefits greatly from local feature learning, which performs an alignment/matching by flipping local features and calculating the shortest path between them. Besides the partial matching method, we also consider a series of data augmentation methods such as flip as new id, random whitening, random crop and so on. And we also use an example sampling strategy, i.e., hard negative mining, for training. In addition, we ensemble the models with different backbones and epochs using imagenet pre-trained models. Extensive experiments validate the superiority of our method for tiger Re-ID. Code has been released at https://github.com/vvictoryuki/tiger_reid_pytorch.