ELViS: Efficient Visual Similarity from Local Descriptors that Generalizes Across Domains

Dimensions overall score 7.0

• ELViS outperforms competing methods by a large margin in out-of-domain scenarios.result 90%
  Evidencepartial

  Our experiments show that ELViS outperforms competing methods by a large margin in out-of-domain scenarios and on average

  Implicationpartial

  Directly stated in abstract with supporting results in Table 1 showing ELViS achieving higher mAP scores compared to other methods on out-of-domain datasets.

  Verificationpartial

  partial

• ELViS requires only a fraction of the computational cost of competing methods.technical 85%
  Evidencepartial

  while requiring only a fraction of their computational cost.

  Implicationpartial

  Directly stated in abstract and supported by Figure 1 showing performance vs. time comparison where ELViS achieves better performance with lower runtime.

  Verificationpartial

  partial

• ELViS operates in similarity space rather than representation space, which promotes cross-domain transfer.method 90%
  Evidencepartial

  Unlike conventional approaches, our model operates in similarity space rather than representation space, promoting cross-domain transfer.

  Implicationpartial

  Explicitly stated in abstract and analysis as a core design principle of the method.

  Verificationpartial

  partial

• Similarity-based models generalize better than descriptor-based ones, which tend to overfit the training domain.result 80%
  Evidencepartial

  Our evaluation confirms that similarity-based models generalize better than descriptor-based ones, which tend to overfit the training domain and excel only on seen distributions.

  Implicationpartial

  Directly stated in analysis section with explanation of why this occurs, though specific comparative evidence is implied rather than explicitly quantified.

  Verificationpartial

  partial

• ELViS refines local descriptor similarities through an optimal transport step with data-dependent gains that suppress uninformative descriptors.method 90%
  Evidencepartial

  It leverages local descriptor correspondences, refines their similarities through an optimal transport step with data-dependent gains that suppress uninformative descriptors

  Implicationpartial

  Detailed description in abstract and Figure 2 shows this as a core technical component of the method.

  Verificationpartial

  partial

• ELViS aggregates strong correspondences via a voting process into an image-level similarity.method 90%
  Evidencepartial

  and aggregates strong correspondences via a voting process into an image-level similarity.

  Implicationpartial

  Explicitly stated in abstract as part of the method description.

  Verificationpartial

  partial

• The paper introduces a benchmark of eight datasets spanning landmarks, artworks, products, and multi-domain collections for evaluating domain generalization.technical 90%
  Evidencepartial

  To assess generalization, we compile a benchmark of eight datasets spanning landmarks, artworks, products, and multi-domain collections

  Implicationpartial

  Explicitly stated in abstract and analysis with specific dataset names listed.

  Verificationpartial

  partial

• ELViS uses a modified BCE loss with a learnable function g that reshapes the penalty curve during training.method 85%
  Evidencepartial

  During training, a modified BCE loss with a learnable function g reshapes the penalty curve

  Implicationpartial

  Detailed description in analysis section and Figure 2 shows this training component, though specific performance impact isn't quantified separately.

  Verificationpartial

  partial

Competitive landscape

A highly efficient image-to-image similarity model that generalizes across diverse domains by operating in similarity space and refining local descriptor correspondences.