Determining traversal space from single image

The model predicts the geometry of both visible and occluded transferable surfaces from the input image. The model may be trained according to a stereoscopic video sequence, using camera gestures, depth per frame, and semantic segmentation to form training data that is used to supervise an image-to-image network. In various embodiments, the model is applied to a single RGB image describing a scene to produce information describing a traversal space including an occluded traversal scene. The information describing the transferable space may include a partition mask of the transferable space (both visible and occluded) and a non-transferable space and a depth map indicating to an estimated depth corresponding to a transferable surface determined to correspond to each pixel of the transferable space. 模型从输入图像预测可见和被遮挡的可穿越表面两者的几何形状。该模型可以根据立体视频序列训练，使用相机姿势、每帧深度和语义分割来形成训练数据，该训练数据用于监督图像到图像网络。在各种实施例中，该模型被应用于描述场景的单个RGB图像以产生描述包括被遮挡的可穿越的场景的可穿越空间的信息。描述可穿越空间的信息可以包括可穿越空间(可见和被遮挡的两者)和不可穿越空间的分割掩码以及深度图，该深度图指示到与可穿越表面对应的估计深度，该可穿越表面