SeqXY2SeqZ: Structure Learning for 3D Shapes by Sequentially Predicting 1D Occupancy Segments from 2D Coordinates

Ray-ONet: Efficient 3D Reconstruction From A Single RGB Image

“…It enables fast mesh extraction but still requires significantly more memory than implicit models. The closest to our work is SeqXY2SeqZ [10], which also tries to reduce the time complexity of inference in implicit representations. To this end, they design a hybrid representation, which descritises the 3D space in two axes while keeping the rest continuous.…”

“…Our work is also a hybrid approach, which explicitly predicts occupancy probabilities along a certain ray that passes through a point on the input image. Unlike Han et al [10], we predict occupancies along a ray in a single network forward, which results in further reduction on time complexity. In addition, the coordinate system of Ray-ONet is aligned with the input image to take advantage of local features.…”

“…Although it is 4× faster than Ray-ONet, it produces inferior reconstruction accuracy and requires much larger memory size with 20× parameters. Since SeqXY2SeqZ [10] has not released their code, we can only compare with them on complexity and accuracy. The inference time complexity of SeqXY2SeqZ is O(N…”

Neural-Pull: Learning Signed Distance Functions from Point Clouds by Learning to Pull Space onto Surfaces

“…Deep learning models have been playing an important role in different 3D computer vision applications [54,49,52,65,27,17,60,4,58,28,3,50,18,23,21,24,25,29,30,20,32,26,22,47,46,31,34,33,19,64]. In the following, we will briefly review work related to learning implicit functions for 3D shapes in different ways.…”

“…Early work requires the ground truth occupancy values or signed distance values as 3D supervision. For single image reconstruction, a single image [63,55,9,41,15,27] or a learnable latent code [54] can be a condition to provide information about a specified shape. For surface reconstruction [66,42,50,15], we can also leverage a point cloud as a condition to learn an implicit function which can be further leveraged to obtain a mesh surface [35,12].…”