AI Paper Atlas

Research map / Point clouds / Segmentation and detection

Group-Free 3D Object Detection via Transformers

Ze Liu, Zheng Zhang, Yue Cao, Han Hu, Xin Tong

ICCV 2021 · Point clouds › Segmentation and detection · approach: Single-stage point-voxel detection · 318 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
use KPS sampling instead of FPS for initial candidates (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 64.5 -> 66.3 mAP@0.25, 46.2 -> 48.5 mAP@0.5Improved
add NMS to KPS sampling (KPS-NMS) (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 66.3 -> 65.8 mAP@0.25, 48.5 -> 48.7 mAP@0.5No clear effect
vary k in KPS sampling (1, 2, 4, 6) (ablation)Group-Free 3D · ScanNet V2mAP@0.25: 65.7, 65.8, 66.3, 66.1 mAP@0.25No clear effect
add iterative spatial-encoding refinement of box location (center+size) (ablation)Group-Free 3D · ScanNet V2mAP@0.5: 43.4 -> 48.5 mAP@0.5, 64.7 -> 66.3 mAP@0.25Improved
add fixed position encoding without iterative refinement (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 64.7 -> 64.6 mAP@0.25, 43.4 -> 43.5 mAP@0.5No clear effect
encode box size in addition to box center in spatial encoding (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 65.2 -> 66.3 mAP@0.25, 47.5 -> 48.5 mAP@0.5Improved
add 6 attention decoder layers (vs 0 layers) (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 63.3 -> 66.3 mAP@0.25, 40.7 -> 48.5 mAP@0.5Improved
add multi-stage ensemble of predictions (ablation)Group-Free 3D · SUN RGB-DmAP@0.25 / mAP@0.5: 59.2 -> 63.0 mAP@0.25, 43.3 -> 45.2 mAP@0.5Improved
add multi-stage ensemble of predictions (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 66.3 -> 66.4 mAP@0.25, 48.5 -> 48.7 mAP@0.5No clear effect
use attention over all points instead of RoI-Pooling grouping (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 65.1 -> 66.3 mAP@0.25, 44.4 -> 48.5 mAP@0.5Improved
use attention over all points instead of Voting grouping (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 64.2 -> 66.3 mAP@0.25, 44.1 -> 48.5 mAP@0.5Improved
widen backbone 2x and use 12 decoder layers (vs 6 layers PointNet++) (ablation)Group-Free 3D · ScanNet V2mAP@0.25 / mAP@0.5: 67.2 -> 68.8 mAP@0.25, 49.7 -> 52.1 mAP@0.5Improved

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