2021-3DV论文阅读 DSP-SLAM-Object Oriented SLAM with Deep Shape Priors

object shapes decoded from latent code

Foreground objects background features and camera poses

monocular stereo sterero+lidar

stereo+LiDAR:

1. incorporates a sparse set of LiDAR measurements (as few as 50 per object) for object reconstruction and pose-only optimization.

2. exactly, 50 3D points per object to obtain accurate shape estimates

for object shape and pose estimation: improve quantitative and qualitatively over auto-labelli

stereo+LiDAR sota

monocular： achieves promising qualitative reconstruction result

对比其他方法：

1. FroDO：batch method
2. Node-SLAM：feature与object没有一起优化
3. DeepSLAM++： forward shape generation不稳定

物体表示：

Each object is represented as a compact and optimizable code vector z

方法：

employ DeepSDF [25] as the shape embedding，输入shape code z（64 dims）和3D query location，outputs the signed distance function (SDF) value s = G(x, z) at the given point

• Detections:

1. 每个关键帧处估计2D bouding box和分割的mask，initial estimate for the object pose estimation来自3D bounding box检测

• Data association:

1. 新的检测关联到 existing map objects，
2. 或者instantiated as a new object via object-level data association
3. 物体实例包括2D bounding box B, a 2D mask M, the dpeth observation of sparse 3D point cloud D, and the initial object pose Tco,0.

• Prior-based object reconstruction:

1. 对于新的实例，输入3D点，optimises the shape code and object pose to minimise surface consistency and depth rendering losses
2. 对于已经存在的实例，只优化 their 6-dof pose

Object Reconstruction with Shape Priors

Surface Consistency Term

Differentiable SDF Renderer

1. 计算Occupancy Probabilities

2. 计算Event Probabilities

3. Rendered Depth and Rendering Term

union of surface pixels and pixels not on object surface but inside the 2D bounding box B

Optimization detail