ICCV Daily 2019 - Tuesday

Qing Liu presenting Localizing Occluders with Compositional Convolutional Networks: they introduce a compositional deep model that is regularized to be fully generative in terms of its high-level features. Their model learns a dictionary of part models, represents 3D objects as mixture of 2D templates, and localizes occluders with high precision. The proposed compositional convolutional network is robust to occlusion and outperforms baseline models on object classification tasks by a large margin. 18 Poster Presentations DA I L Y YimingWang presenting Active 3D Classification of Multiple Objects in Cluttered Scenes. Robust object recognition is an essential skill for robots to perform autonomous manipulation. While classifying objects in cluttered scenes can be challenging with single-shot methods due to severe occlusion. They therefore propose an active vision approach through 3D reconstruction following a next- best-view (NBV) paradigm.