Learning-Deep-Learning

OpenOccupancy: A Large Scale Benchmark for Surrounding Semantic Occupancy Perception

March 2023

tl;dr: A new benchmark of NuScenes-Occupancy, with new KPI and simple baselines.

Overall impression

The paper used an Augmenting And Purifying (AAP) pipeline for autolabel. This essentially leverages a functional (yet not performing) pretrained baseline model to prelabel. Human labelers are then asked to purify the pre-label. This way it can greatly boost labeling efficiency.

This pipeline aims to provide dense occupancy labels, with semantic information. It invests 4000 human hours to refine the label. This roughly translates to 4 human hours per 20-s clip.

The idea of Cascade Occupancy Network (CONet) to improve latency is relevant to industrial deployment. The simple idea of bilinear interpolation seems to work quite well already.

The pipeline to generate occupancy label is Poisson Recon used by SurroundOcc, 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.

Key ideas

Technical details

Notes