For this assigment, I have tried different distances for histogram measurements, using different p values in Lp metrics, as well as cos distance. The results are generally poor. For each query picture, among the top 5 most similar images, only 2 or 3 are correct (the same class with the query picture). For some cases, even no correct image is picked up. I think the reason is: 1) feature is few. Only one feature is used in retrieval, 2) feature is not representative. Maybe histogram can not reflect essential characters of the images, and 3) information is lossed when using this feature. The spatial information is lossed when using histogram as the feature of images.