Discovering Object Instances from Scenes of Daily Living

1. Introduction


We tackle the problem of discovering object instances from Activities of Daily Living. Imagine a personal robotic assistant that accompanies a user to different scenes during her daily activities, such as kitchen, living room and office space. Every time the user operates in a scene, the robotic assistant takes a few pictures of the environment without the user explicitly showing the objects to the robot (a). After a period of data gathering, we would like the robotic assistant to automatically discover and model objects from the images it has collected (b).

2. Examples of discovered objects

We evaluate our program on three different datasets, each with different characteristics.

ADL dataset (heterogenity of object appearance, cluttered background, occlusion and viewpoint change)

CMU object dataset (Cluttered environment, occlusion and viewpoint changes)

Flickr images (Interference from irrelevant images)