AI Paper Atlas

Research map / Detection and segmentation / CNN features

Parameter-Inverted Image Pyramid Networks for Visual Perception and Multimodal Understanding

Zhaokai Wang, Xizhou Zhu, Xue Yang, Gen Luo, Hao Li, Changyao Tian, Wenhan Dou, Junqi Ge, Lewei Lu, Yu Qiao, Jifeng Dai

TPAMI 2025 · Detection and segmentation › CNN features · approach: Convolution-transformer backbones

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
replace ViTDet-B by parameter-inverted pyramid PIIP-TSB at similar FLOPsViTDet-B · COCO Mask R-CNN 1xbox AP: 43.8 -> 46.6Improved
replace ViTDet-B by PIIP-TSB at lower FLOPs (243G vs. 463G)ViTDet-B · COCO Mask R-CNN 1xbox AP / FLOPs: 43.8 -> 43.9 with 463G -> 243GImproved
replace ViTDet-L by PIIP-SBL at similar FLOPsViTDet-L · COCO Mask R-CNN 1xbox AP: 46.8 -> 49.4Improved
replace InternViT-6B by PIIP-LH6BInternViT-6B · COCO Mask R-CNN 1xbox AP: 53.8 -> 55.7Improved
replace InternViT-6B by PIIP-LH6BInternViT-6B · ADE20K UperNetmIoU: 58.36 -> 59.65Improved
use image pyramid via multi-scale training on a single branch (ablation)ViTDet-B · COCO Mask R-CNN 1xbox AP: 43.8 -> 44.8Improved
use parameter-inverted resolution order instead of parameter-direct (ablation)PIIP-TSB · COCO Mask R-CNN 1xbox AP: 42.6 -> 46.6Improved
use ConvNeXt for the high-resolution branch and ViT for lower ones instead of all ViT (ablation)PIIP-TSB · COCO Mask R-CNN 1xbox AP: 44.3 -> 47.1Improved
merge outputs of all three branches instead of only branch B (ablation)PIIP-TSB 1568/896/672 · COCObox AP: 43.1 -> 46.6Improved
use deformable attention instead of regular attention (12 interactions) (ablation)PIIP-TSB 1120/896/448 · COCObox AP: 42.0 -> 43.9Improved
increase interactions from 0 to 12 with deformable attention (ablation)PIIP-TSB 1120/896/448 · COCObox AP: 41.3 -> 43.9Improved
add 2 interactions to the vision encoder (ablation)PIIP-LLaVA CLIP-B+CLIP-L · multimodal 8-benchmark averageAvg: 64.4 -> 65.2Improved

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