Name of Reviewer ------------------ Xu Miao Learning Visual Similarity Measures for Comparing Never Seen Objects Key Contribution ------------------ Intuitively positive and negative examples should have different mass densities on a common feature space. They come up with a nice feature space and train a set of extremely randomized trees to partition the feature spaces. Surprisingly the densities characterized based on that partitions can be linearly separated with low confusion. Novelty -------- At least the feature space representation is new to me. Reference to prior work ----------------------- Clarity ------- The motivation of the approach from the application problem (determining if a pair of never-seen object is "same" or "different") is stated well. But it lacks an overview of the difficuties on solving this problem and a methodolgy comparison to other approaches, or simply what is state-of-arts methods applied in this problem. There is some confussion in the paper. For example, why they claim the weight is low if a leaf is equally probable in pos and neg image pairs and .... But the weight seems to be deteremined by a SVM which maximizes the margin between the data not concerning about the densities. They didn't elaborate it in any depth or in experiments. Moreover, I am just a little bit surprised the positive and negative examples can be simply separated by a hyper-plane. It might be because of the dimensionality of x is high enough, but the authors didn't make any comments. Technical Correctness --------------------- The paper looks technically sound. Experimental Validation ----------------------- The experiments are good. They also did an experiment to show that the model is category-dependent. This weaken the title of comparing never seen objects. It might be better to title "Learning visual similarity measures for comparing never seen objects in the same category". Overall Evaluation ------------------ It does prove the feature space is quite useful. The experiments are clean. Questions and Issues for Discussion ----------------------------------- I wonder why not split the node until every leaf contains only positive or negative image pairs. In this case, the code of a positive example is gauranteed to be different than one of a negative example. What makes it possible to get a category-independent model which can comparing truely never seen objects?