AI Paper Atlas

Research map / 3D vision / 3D generation

Scene as Occupancy

Wenwen Tong, Chonghao Sima, Tai Wang, Li Chen, Silei Wu, Hanming Deng, Yi Gu, Lewei Lu, Ping Luo, Dahua Lin, Hongyang Li

ICCV 2023 · 3D vision › 3D generation · approach: Vision-based 3D occupancy prediction · 153 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
decode voxel features from BEV features with a cascade voxel decoder instead of BEVNet (ablation)OccNet · OpenOcc val (ResNet50)mIoU: 17.37 -> 19.48Improved
use cascade voxel decoder instead of direct VoxelNet (ablation)OccNet · OpenOcc val (ResNet50)mIoU: 19.06 -> 19.48; FPS 1.9 -> 2.6; memory 23G -> 18GImproved
add occupancy-based scene completion features as perception input for planning rasterization instead of bounding boxesST-P3-style planner · nuScenes planningCollision rate 1s (%): 0.27 -> 0.21Improved
use all 16 occupancy classes for planning (ablation)OccNet planning · nuScenes planningL2 3s (m): L2 3s 3.07 -> 2.99; collision 2s 0.55 -> 0.59Mixed
jointly train scene completion with 3D detection (ablation)BEVNet, VoxNet, OccNet · nuScenes val 3D detectionmAP / NDS: BEVNet mAP 0.259 -> 0.271; OccNet NDS 0.382 -> 0.390Improved
joint training with occupancy slightly worsens box translation and scale errors (ablation)BEVNet · nuScenes val 3D detectionmATE: 0.828 -> 0.835Hurt
pretrain on occupancy instead of detection for BEV segmentation fine-tuningBEV segmentation · nuScenes BEV segmentationIoU main value: 18.18 -> 19.17Improved
pretrain on occupancy rather than detection for planningST-P3 planner · nuScenes planningCollision rate 3s (%): 0.82 -> 1.03Hurt
use finer voxel size 0.25m instead of 0.5m for irregular objects (ablation)Occupancy vs detection · nuScenes irregular objectsmIoU_ir gap: +2.21 -> +3.51Improved
increase history frames for temporal self-attention from 2 to 4 (ablation)OccNet · OpenOcc val (ResNet50)mIoU: 19.48 -> 20.81Improved
use only one history frame (ablation)OccNet · OpenOcc val (ResNet50)mIoU: 19.21 (0 frames) -> 18.35 (1 frame)Hurt
decrease voxel size for LiDAR segmentation transfer (ablation)OccNet · nuScenes val LiDAR segmentationmIoU: 46.60 (1.0m) -> 47.29 (0.5m) -> 53.00 (0.25m)Improved

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy