Object discovery in depth images

We present an unsupervised method for discovering objects from depth information. Our method can identify new common objects appearing in different depth images. We use 2D bounding box proposals to detect candidate locations of objects in each depth image, and then retrieve the corresponding 3D bounding boxes using the depth information. Invalid object proposals can be further removed by analyzing the point cloud distribution inside the 3D bounding box. We measure the similarity between each pair of the object proposals in different images to identify co-occurrences of the same instance. The similarity measure is automatically learned by a Siamese convolutional neural network. Our method is unsupervised in a sense that we do not need human labeled data to train the Siamese network. We use 3D CAD models to synthesize a large set of similar and dissimilar pairs of depth images as the positive and negative data. Our experiments on synthetic data show that the proposed method is able to discover the co-occurrences of the common objects in different depth images.