Title: Top-5 problems in multimedia retrieval
June 14, 9:30-10:30, Hall
- Tat-Seng Chua, National University of Singapore
- Michael Houle, National Institute of Informatics
- Ramesh Jain, University of California, Irvine
- Nicu Sebe, University of Trento
- Rainer Lienhart, University of Augsburg
- Chong-Wah Ngo, City University of Hong Kong
- Vincent Oria, New Jersey Institute of Technology
Multimedia retrieval is hard, but what exactly are the core problems that are fundamentally important for all/most sub-fields of retrieval? Semantic gap is well regarded as a core problem since it was first named in year 2000. But after almost 20 years, how much have we tackled the problem? In addition to semantic gap, we have user gap. Is user gap a core problem? If so, why relatively more papers in this field are about semantic gap rather than user gap? Fusion should be a fundamental problem because we deal with different media, multiple modalities and multi-sensory data. But, most of the published papers use average or max fusion (pooling) or simply let neural network learn an “embedding space” to fuse different forms of data. Is fusing data in this ad-hoc manner sufficient and scientific in the context of retrieval? This community is particularly good in dealing with big data. But since the advancement of deep learning, more researches are about using big training data but small testing data for retrieval and annotation. Scalability is no longer an issue? Retrieval is supposed to be real-time and requires indexing. Is indexing multimedia data a core problem, or is it sufficient to just index individual media using off-the-shelf techniques and then combine the result for retrieval? Another issue that is seldom discussed in this community is the effect of the so called “curse of dimensionality” on Multimedia information retrieval. Multimedia data are represented as vectors in high dimensional spaces. As the dimensionality increases, the discriminative ability of similarity measures diminishes to the point where methods such as search and clustering that depend on them lose their effectiveness.
Are these problems (semantic gap, user gap, fusion, scalability, indexing) fundamentally important to retrieval? How far have we reached the solutions that the community can endorse? How do the current solutions impact the future era of retrieval and the applicability of multimedia search in different applications? This panel aims to solicit expert views from senior researchers who have actively contributed to the field for more than 20 years, and possibly pave a way for new research direction(s) in multimedia retrieval.