In the context of smart cities, one fundamental visual task is to effectively and efficiently detect people by means of inferring their spatial locations on the image frame, so that insights for urban surveillance can be provided to law enforcement agencies. However object detection technologies often suffer from performance drop when the model is trained from a data domain that is different from the data domain it should be tested. Fine-tuning the model on the target domain can often improve the performance, but at the pain of manual annotations. Our unit develops not only easy-to-use toolsets to ease the manual annotation, but also focuses on domain adaptation strategies for object detector.Â