CVPR2023-3D-Occupancy-Prediction

Why we included this project

This repository is the official resource for the CVPR 2023 3D Occupancy Prediction Challenge, organized around the idea that autonomous driving perception needs the whole scene, not just boxes. Classic 3D detection describes foreground objects with bounding boxes, which works for vehicles and pedestrians but flattens odd-shaped obstacles and ignores the background entirely. This task instead predicts a voxelized occupancy map built on nuScenes, where each 0.4-meter voxel records whether the space is free or occupied and, if occupied, which of 18 semantic labels it carries. The repo does not ship a deployable system; it gives you a working baseline, a devkit to parse and score results, and the same mIoU and F-score metrics used to rank challenge entries. That is a useful common yardstick when you want to compare your surround-view occupancy model against the published leaderboard, or when you want to see exactly how the benchmark ground truth was put together before building your own dense perception approach.