Papers › The Missing Point in Vision Transformers for Universal Image Segmentation

The Missing Point in Vision Transformers for Universal Image Segmentation

26 May 2025arXiv:2505.19795archive 2025-07-28

Sajjad Shahabodini, Mobina Mansoori, Farnoush Bayatmakou, Jamshid Abouei, Konstantinos N. Plataniotis, Arash Mohammadi

Image segmentation remains a challenging task in computer vision, demanding robust mask generation and precise classification. Recent mask-based approaches yield high-quality masks by capturing global context. However, accurately classifying these masks, especially in the presence of ambiguous boundaries and imbalanced class distributions, remains an open challenge. In this work, we introduce ViT-P, a novel two-stage segmentation framework that decouples mask generation from classification. The first stage employs a proposal generator to produce class-agnostic mask proposals, while the second stage utilizes a point-based classification model built on the Vision Transformer (ViT) to refine predictions by focusing on mask central points. ViT-P serves as a pre-training-free adapter, allowing the integration of various pre-trained vision transformers without modifying their architecture, ensuring adaptability to dense prediction tasks. Furthermore, we demonstrate that coarse and bounding box annotations can effectively enhance classification without requiring additional training on fine annotation datasets, reducing annotation costs while maintaining strong performance. Extensive experiments across COCO, ADE20K, and Cityscapes datasets validate the effectiveness of ViT-P, achieving state-of-the-art results with 54.0 PQ on ADE20K panoptic segmentation, 87.4 mIoU on Cityscapes semantic segmentation, and 63.6 mIoU on ADE20K semantic segmentation. The code and pretrained models are available at: https://github.com/sajjad-sh33/ViT-P}{https://github.com/sajjad-sh33/ViT-P.

PaperPDFCode

Code

sajjad-sh33/vit-p mentioned in paperpytorch report

Repository list and official/mentioned flags are the archive's, frozen 2025-07-28. Reachability, where shown, is from one Syntology probe window (2026-09-16 to 2026-09-18); repositories not probed show nothing. GitHub stars are not tracked.

Code Syntology ran Syntology

Not run by Syntology. Nothing on this page verifies that the listed code works.

Tasks

Image SegmentationInstance SegmentationPanoptic SegmentationSegmentationSemantic SegmentationUniversal Segmentation

Results from the paper archive 2025-07-28

TaskDatasetModelMetricValueRank at snapshotLeaderboardReport
Instance Segmentation ADE20K val ViT-P (OneFormer, DiNAT-L, single-scale, 1280x1280, COCO_pretrain) AP 40.7 #3 of 14 Archive leaderboard report
Instance Segmentation ADE20K val ViT-P (OneFormer, DiNAT-L, single-scale, 1280x1280) AP 37.8 #6 of 14 Archive leaderboard report
Instance Segmentation Cityscapes val ViT-P (OneFormer, ConvNeXt-L, single-scale, 512x1024, Mapillary Vistas-pretrained) AP 49.0 #1 of 17 Archive leaderboard report
Instance Segmentation Cityscapes val ViT-P (OneFormer, ConvNeXt-L, single-scale, 512x1024, Mapillary Vistas-pretrained) mask AP 49.0 #1 of 17 Archive leaderboard report
Panoptic Segmentation ADE20K val ViT-P (OneFormer, DiNAT-L, single-scale, 1280x1280, COCO_pretrain) PQ 54.0 #2 of 25 Archive leaderboard report
Panoptic Segmentation ADE20K val ViT-P (OneFormer, DiNAT-L, single-scale, 1280x1280) PQ 51.9 #7 of 25 Archive leaderboard report
Panoptic Segmentation Cityscapes val ViT-P (OneFormer, InternImage-H) AP 50.6 #1 of 37 Archive leaderboard report
Panoptic Segmentation Cityscapes val ViT-P (OneFormer, InternImage-H) PQ 70.8 #1 of 37 Archive leaderboard report
Panoptic Segmentation Cityscapes val ViT-P (OneFormer, InternImage-H) mIoU 85.4 #1 of 37 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (InternImage-H) Params (M) 1610 #1 of 235 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (InternImage-H) Validation mIoU 63.6 #1 of 235 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (OneFormer, InternImage-H) Params (M) 1400 #7 of 235 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (OneFormer, InternImage-H) Validation mIoU 61.6 #7 of 235 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (OneFormer, DiNAT-L) Params (M) 309 #14 of 235 Archive leaderboard report
Semantic Segmentation ADE20K ViT-P (OneFormer, DiNAT-L) Validation mIoU 59.9 #14 of 235 Archive leaderboard report
Semantic Segmentation COCO (Common Objects in Context) ViT-P (OneFormer, InternImage-H) mIoU 69.1 #2 of 9 Archive leaderboard report
Semantic Segmentation COCO (Common Objects in Context) ViT-P (OneFormer, DiNAT-L) mIoU 68.8 #4 of 9 Archive leaderboard report
Semantic Segmentation COCO-Stuff test ViT-P (InternImage-H) mIoU 53.5 #2 of 21 Archive leaderboard report
Semantic Segmentation Cityscapes val ViT-P (InternImage-H) mIoU 87.4 #1 of 99 Archive leaderboard report

Ranks are positions in the archive's leaderboards as they stood at the 2025-07-28 snapshot. Results published since then are not among these rows, so a rank here is not a current standing.

Methods

Absolute Position EncodingsAdamAttentionBPEDense ConnectionsDropoutLabel SmoothingLayer NormalizationLinear LayerMulti-Head AttentionPosition-Wise Feed-Forward LayerResidual ConnectionSoftmaxTransformerVision Transformer

Report a problem or propose a change · a person checks every report against the paper or source before anything changes; decisions are listed on /corrections