| Summary: | 碩士 === 國立成功大學 === 資訊工程學系 === 104 === In the current work, we have proposed an automatic 2D-to-3D depth estimation method. It is primarily based on learning and geometry. Because it is a learning-based algorithm, it has no restriction on the scene of the input image or video. The new method can also refine the depth estimation result by vanishing point in geometry. In order to accelerate the computing speed of the overall algorithm, we extract the features of 2D images in the database in advance. Next, to find the most similar images within the database by image features, the only necessary thing to do is to compute the feature of the query image and compare the difference between the query image and training images. Such an approach can reduce computing time. Furthermore, the paper also introduces the method of propagating depth value to video with similar scenes. The experimental results show that the depth map of our algorithm is not only closely similar to ground truth, but also the PSNR and VIF (visual information fidelity) show better performance than other reported algorithms.
|
|---|
