Data collection sampling
This work was published here .
In our previous work, we proposed a model to measure the amount of information of a dataset based on the volume spanned by their description in a semantic space. Let us consider that each image is described in a semantic space by a vector z . The image descriptions are concatenated in a matrix Z . The dataset information is modeled as the log-determinant of the pseudo covariance matrix: log(| ZZ T |). The sampling can be done by selecting the columns of Z , such that they maximize the metric above. In this work we have also introduced the notion of user’s preference so that one can control which information is more important to the user. We also propose an algorithm for volume maximization.
Exploiting Inter-item correlation
In a large dataset, the different images may contain some common information that does not necessarily reside in the pixel space. For example, the images may contain different types of landscape, animals, that could look different but still describe the same category of objects. We have explored solutions to exploit such correlations to reduce the compressed dataset size.
- Semantic alignment (-PDF-) where we shape the latent space of an auto-encoder to align the semantically coherent image together
- SMIC (-PDF-): a dictionary-based compression where we exploit the linearity of CLIP representation