Automatic musical instrument recognition is a relatively new topic in the growing field of Music Information Retrieval. Early studies mostly focused on instrument recognition from recordings of isolated notes. More recently, some studies tackled the problem of musical phrases played in solo (i.e. without accompaniment) which better covers the timbre variability of a given instrument. However, the current trend is now to deal with true polyphonic music (i.e. involving multiple instruments), which appears to be a far more difficult problem but with more practical applications. The aim of this paper is to provide an overview of the state-of-the-art in automatic musical instrument recognition with a focus on recent and innovative approaches applied to true polyphonic music. It will be shown that the traditional "bag of frames" approaches can obtain interesting results by building efficient automatic taxonomies or by using complementary information to enhance the relevant signal. We however argue that it is important to consider new directions to overcome the limitations of these traditional approaches. One of these promising directions that will be detailed concerns mid-level representations, which are based on the decomposition of the signal into a small number of sound atoms or molecules bearing explicit musical instrument labels. INTRODUCTION There is a growing interest for new means of interaction with audio information that is nowadays mostly available in digital format and stored in large databases. There is therefore a strong need for efficient audio indexing techniques that would allow the extraction of a detailed and meaningful symbolic representation directly from a digital audio recording. For music signals, this representation will include information about the metric, the harmony, the melody, the genre or the interpretation style and will ultimately be represented under the form of an enriched music sheet. The availability of such a symbolic representation opens the path for numerous Music Information Retrieval (MIR) applications including content-based search by similarity, cover songs retrieval, automatic post-remixing,. .. .