In Social Signal Processing (SSP) and affective computing area, the publicly available facial expression dataset for emotion recognition task is still limited for basic emotion categories. Whereas in everyday life, various types of emotions are being used by humans more than basic emotion, such as mixed emotion. To enrich the diversity of the existing dataset, we developed the Indonesian Mixed Emotion Dataset (IMED). The objective of creating this dataset is to provide the annotated data for mixed emotion recognition as a ground-truth for benchmarking. Mixed emotion is constructed by combining basic emotion categories to resulting new ones. This dataset can be used to facilitate mixed emotion recognition experiments. Our dataset displays 19 categories of emotions performed by 15 subjects, all are Indonesians with various ethnicities: Javanese, Sundanese, Malay, Bataknese, Minang, and Manadonian. Subjects are 60% female and 40% male with age ranging from 17 to 32 demonstrated basic and mixed emotion classes in videos. We then used a computational model to show that mixed emotion categories were discriminable to be recognized by machine classifiers. We believe that IMED dataset is useful for researchers on the same field to test their novel method by using our dataset.