ReID via Ranking Aggregation of Similarity Pulling and Dissimilarity Pushing

关注 ranking optimization

this paper considers both similarity and dissimilarity relationships for ranking optimization in an automatic manner to further improve person re-identification performance.

conductive similarity：if a gallery is strongly similar to the friends of the probe, it will be much likely to be a friend of the probe.
insulative dissimilarity: if a gallery is very similar to the strangers of the probe, it is much prone to differ from the probe.

propose a ranking aggregation method combining two different person reidentification methods to enhance the conductive similarity and insulative dissimilarity.

Ranking aggregation

• score-based re-ranking
• order-based re-ranking

作者考虑基于order的re-ranking：1、对异常值鲁棒 2、基于score能转换为基于order的

• similarity ranking aggregation：improve the ranking orders of quasi-similar galleries
• dissimilarity ranking aggregation：penalizes the quasi-dissimilar galleries

Framework

1. choose two different baseline methods with more complementarities
2. similarity and dissimilarity ranking aggregation
3. combine（SRA修正原始ranking，然后用DSRA修正SRA的ranking）

Similarity Ranking Aggregation

1. 从两个原始方法的top-k中选出交集作为强相似的gallery
2. 把强相似的gallery做为probe再次查询（为了增加互补性，采用cross-view based backward requery用一种方式选出来的强相似gallery，用另一种方式去requery）
3. graph-based weighted reranking
   对于2：
   
   
   where | · | denotes the cardinality and w(p, g+) is a weighting coefficient related to the original rank in G+(p). The decay factor is defined as w0 and we set w0 = 0.8 the same as [39] in all experiments.
   然后用crossed backward requery求sim2，然后合成：
   
   然后用这个相似度 rerank强相似的gallery，其他的用原始的相似度（α * RL1 + （1-α）*RL2）

Dissimilarity Ranking Aggregation

1. 从两个原始方法的last-k中选出并集作为强不相似的gallery
2. 对于每一个强不相似的gallery求其两种方法的top-k，然后求并集
3. 
4. 用这个频率去修正之前SMA中求出的rerank