*May 2023*

tl;dr: Dense annotation generation for 3D Occupancy Prediction.

Occupancy grid can describe real-world objects of arbitrary shapes and infinite classes. SurroundOcc proposed a pipeline to generate dense occupancy GT without expansive occupancy annotation by human labelers. The paper also demonstrate very clearly that with denser label, previous method (TPVFormer)’s performance can be significantly boosted, almost by 3x. This is the largest contribution of this paper.

The paper is from the same group of SurroundDepth, which performs bottom-up depth estimation from the standpoint of the **source**. In comparison, SurroundOcc performs occupancy prediction from the standpoint of the **target**. This relationship is quite similar to that between Lift-Splat-Shoot and BEVFormer.

The pipeline to generate occupancy label is Poisson Recon used by SurroundDepth, and Augment-and-Purify (AAP) pipeline proposed by OpenOccupancy. They share the initial steps but are different in the refinement step. It would be interesting to see a side-by-side comparison of the two.

- Architecture largely recycles BEVformer with deformable attention.
- Automatic annotation of dense occupancy labels
- To get dense predictions, we need dense occupancy labels.
- By stitching multi-0frame lidar points of dynamic objects and static scenes separately. P = [P_s, P_o]
- Densification with Poisson Reconstruction
- P is already denser than single frame measurement, but still has many holes. Reconstruct P to a triangular mesh M via
**Poisson Surface Reconstruction**. - Poisson Surface Recon: input is point cloud with normal vectors, output is triangular mesh.
- M = {V, E}, where vertices V is evenly distributed, then we can convert this to dense voxels Vd. Compared with Vs the direct voxelization from P, Vd is much denser.
- For each voxel in Vd, then use
**NN (nearest neighbor)**to assign the semantic labels in Vs.

- cross-entropy loss and scene-class affinity loss as proposed in MonoScene. High resolution prediction is more important, a decayed loss weight 1/2^j for j-th level supervision.

- Most 3D scene reconstruction (Atlas, NeuralRecon, TransformerFusion) methods are focused on indoor scenes.
- CD is accuracy and complementary accuracy average, and thus is permutation invariant. CD and F1 score both consider precision and recall. These two metrics are typical metrics for indoor scene reconstruction.
- Range is [-50m, 50m] x [-50m, 50m] x [-5m, 3m]. 200x200x16 with 0.5m voxel size.
- TSDF fusion fuses depth measurements into coherent point clouds.