Figure 1 - uploaded by Laksmita Rahadianti
Content may be subject to copyright.
Source publication
Distance metric learning has been considered an effective strategy to represent data in computer vision problems such as image retrieval and face verification. Metric learning attempts to minimize a loss function in order to transform data into a more optimal representation for further applications. In this paper, we compare 4 different types of lo...
Contexts in source publication
Context 1
... Loss. The two Convolution Neural Networks, as illustrated in Fig.1 are not different but are two copies of the same network, hence the name Siamese Networks [4]. ...
Context 2
... for the vector bank. Our models were requested to retrieve the face images that are most similar to the query from the vector bank. Previously we noticed that image embedding based on Contrastive loss failed to push samples that were in different classes. This indeed was apparent in the retrieval result. For example, the result as illustrated in Fig. 10 shows that model meets difficulty in distinct images that have identical features. In some instances, the model can distinguish between gender and hair style well but is unsuccessful to differentiate a specific person. Next we show the image retrieval results using the Triplet Margin Ranking loss in Fig. 11). Even though it performs ...
Context 3
... example, the result as illustrated in Fig. 10 shows that model meets difficulty in distinct images that have identical features. In some instances, the model can distinguish between gender and hair style well but is unsuccessful to differentiate a specific person. Next we show the image retrieval results using the Triplet Margin Ranking loss in Fig. 11). Even though it performs better in the image retrieval task but it also had difficulties to retrieve specific individuals from the LFW dataset. Observing the high dimensional visualization for face embedding using Triplet margin ranking loss, several classes were still relatively close to each other. This may have happened due to the ...
Context 4
... margin ranking loss, several classes were still relatively close to each other. This may have happened due to the output vector size being too small to represent large number of unique labels. Finally, the face embedding model based on Proxy-NCA loss were able to give a satisfactory result. Looking at the image retrieval results as illustrated in Fig. 12, Triplet Margin Ranking loss, and Proxy-NCA loss relatively gave almost the same retrieval results. Finally, we also performed a face image retrieval task using the proxy anchor loss. Combining the advantage of data-to-data relation and proxy as an anchor, this loss performed better than other losses based on image retrieval result ...
Context 5
... Triplet Margin Ranking loss, and Proxy-NCA loss relatively gave almost the same retrieval results. Finally, we also performed a face image retrieval task using the proxy anchor loss. Combining the advantage of data-to-data relation and proxy as an anchor, this loss performed better than other losses based on image retrieval result that shown in Fig. 13. ...
Similar publications
In recent years, the face recognition algorithm based on CNN mainly focuses on the improvement of loss function. This paper improves the softmax loss function from two aspects: one is to normalize the weight and features of the loss function, the other is to introduce in class cosine similarity judgment on the basis of the original loss function. T...
