AI Paper Atlas

Research map / Point clouds / Segmentation and detection

ProposalContrast: Unsupervised Pre-training for LiDAR-Based 3D Object Detection

Junbo Yin, Dingfu Zhou, Liangjun Zhang, Jin Fang, Cheng-Zhong Xu, Jianbing Shen, Wenguan Wang

ECCV 2022 · Point clouds › Segmentation and detection · approach: LiDAR voxel and pillar detectors · 110 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
add attentive proposal encoder instead of max pooling (ablation)ProposalContrast, CenterPoint · Waymo val 20%mAPH: 59.32 -> 60.13 (+0.81 mAPH)Improved
pre-train with proposal-level contrast instead of random init (ablation)CenterPoint VoxelNet · Waymo val 20%mAPH: 56.96 -> 60.13 (+3.17)Improved
pre-train at proposal level rather than point level (PointContrast) (ablation)CenterPoint VoxelNet · Waymo val 20%mAPH: 57.75 -> 60.13Improved
pre-train at proposal level rather than scene level (DepthContrast) (ablation)CenterPoint VoxelNet · Waymo val 20%mAPH: 57.96 -> 60.13Improved
add inter-cluster separation (ICS) to inter-proposal discrimination (IPD) (ablation)ProposalContrast · Waymo val 20%mAPH: 59.56 -> 60.13Improved
add Sinkhorn-Knopp clustering to ICS (ablation)ICS-only pre-training · Waymo val 20%mAPH: 54.52 -> 59.16 (-2.44 mAPH without SKC relative to baseline)Improved
use 2048 proposals instead of 1024 (ablation)ProposalContrast · Waymo val 20%mAPH: 59.47 -> 60.13Improved
use 2048 proposals instead of 4096 (ablation)ProposalContrast · Waymo val 20%mAPH: 59.91 -> 60.13Improved
use radius 1.0 m instead of 0.5 m or 2.0 m (ablation)ProposalContrast · Waymo val 20%mAPH: 59.81 (0.5) / 59.55 (2.0) -> 60.13 (1.0)Improved
use 128 clusters instead of 64 or 256 in ICS (ablation)ICS-only pre-training · Waymo val 20%mAPH: 58.87 (64), 59.16 (128), 58.91 (256)Mixed
pre-train with ProposalContrast on small labeled fraction (1%)CenterPoint VoxelNet · Waymo val 1% labelsAPH: 17.83 -> 34.78Improved
pre-train with ProposalContrast when labels are plentiful (100%)PointPillars / VoxelNet · Waymo val 100% labelsAPH: +0.40 / +0.61 APHNo clear effect

Later papers that compared with this method

3 later accepted papers report ProposalContrast as a baseline in their results tables; 2 of them report a better number on most of the results they share with it.

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