This paper presents a novel approach for instance search and object detection, applied to museum visits. This approach relies on fully convo-lutional networks (FCN) to obtain region proposals and object representation. Our proposal consists in four steps: a classical convolutional network is first fined-tuned as classifier over the dataset, next we build from this network a second one, fully convolutional, trained as classifier, that focuses on all regions of the corpus images, this network is used in a third step to define image global descriptors in a siamese architecture using triplets of images, and eventually these descriptors are then used for retrieval using classical scalar product between vectors. Our framework has the following features: i) it is well suited for small datasets with low objects variability as we use transfer learning, ii) it does not require any additional component in the network as we rely on classical (i.e. not fully convolutional) and fully convolutional networks, and iii) it does not need region annotations in the dataset as it deals with regions in a unsupervised way. Through multiple experiments on two image datasets taken from museum visits , we detail the effect of each parameter, and we show that the descriptors obtained using our proposed network outperform those from previous state-of-the-art approaches.