Unsupervised Multi-View CNN for Salient View Selection and 3D Interest Point Detection

We present an unsupervised 3D deep learning framework based on a ubiquitously true proposition named by us view-object consistency as it states that object and its projected 2D views always belong to the same class. To validate effectiveness, we design multi-view CNN instantiating for salient view selection interest point detection of objects, which quintessentially cannot be handled supervised due difficulty collecting sufficient consistent training data. Our CNN, namely UMVCNN, branches off two channels encode knowledge within each respectively also exploits both intra-view inter-view object. It ends with new loss layer formulates impelling generate classification outcomes. The UMVCNN is then integrated global distinction adjustment scheme incorporate cues into selection. evaluate our method section qualitatively quantitatively, demonstrating superiority over several state-of-the-art methods. In addition, showcase can used select scenes containing multiple objects. develop conduct comparative evaluations publicly available benchmark, shows amenable different shape understanding tasks.